Understanding digital signal processing 9789811049613, 9789811049620

Citation preview

Springer Topics in Signal Processing

Orhan Gazi

Understanding Digital Signal Processing

Springer Topics in Signal Processing Volume 13

Series editors Jacob Benesty, Montreal, Canada Walter Kellermann, Erlangen, Germany

More information about this series at http://www.springer.com/series/8109

Orhan Gazi

Understanding Digital Signal Processing

123

Orhan Gazi Electronics and Communication Engineering Department Çankaya University Etimesgut/Ankara Turkey

ISSN 1866-2609 ISSN 1866-2617 (electronic) Springer Topics in Signal Processing ISBN 978-981-10-4961-3 ISBN 978-981-10-4962-0 (eBook) DOI 10.1007/978-981-10-4962-0 Library of Congress Control Number: 2017940604 © Springer Nature Singapore Pte Ltd. 2018 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Printed on acid-free paper This Springer imprint is published by Springer Nature The registered company is Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore

Preface

In this book, we tried to explain digital signal processing topics in detail. We paid attention to the simplicity of the explanation language. And we provided examples with increasing difficulty. The reader of this book should have some background about signals. If it is possible, the reader should learn fundamental concepts on signals and systems since, in this book, more attention is paid on digital signal processing concepts rather than continuous time signal processing topics. Hence, we assume that the reader has fundamental knowledge about all types of signals and transforms. All the topics in this book are presented in an orderly manner. We tried to simplify the language of this book as possible as we can. We also provided original examples explaining the aim of the subjects studied in this book. Numerical examples are provided for the comprehension of the subjects. Unnecessary abundance of mathematical details is omitted for the simplicity of the presentation language. In addition, to indicate both continuous and digital time frequencies, we preferred to use the same parameter. We thought that using two different parameters mixes the students’ mind and it is not necessarily needed. This book includes four different chapters. And in these chapters, sampling of continuous time signals, multirate signal processing, discrete Fourier transform, and filter design concepts are covered. In sampling of continuous time signals and multirate signal processing chapters, we provided some original practical techniques to draw the spectrum of aliased signals. In discrete time Fourier transform chapter, well-designed numerical examples are provided to illustrate the operation of the fast Fourier transform algorithm. In filter design chapter, both analog and digital filter design techniques are explained in detail. For the analog filters, we also provided analog filter circuit design methods for the designed analog filter transfer function. Maltepe/Ankara, Turkey November 2016

Orhan Gazi

v

Contents

1 Sampling of Continuous Time Signals . . . . . . . . . . . . . . . . . . . . . 1.1 Sampling Operation for Continuous Time Signals . . . . . . . . . 1.1.1 Sampling Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.2 Mathematical Characterization of the Sampling Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Sampling Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.1 The Fourier Transform of the Product Signal . . . . . . . 1.3 How to Draw Fourier Transforms of Product Signal and Digital Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 Drawing the Fourier Transform of Digital Signal . . . . 1.4 Aliasing (Spectral Overlapping) . . . . . . . . . . . . . . . . . . . . . . . 1.4.1 The Meaning of the Aliasing (Overlapping) . . . . . . . . 1.4.2 Drawing the Frequency Response of Digital Signal in Case of Aliasing (Practical Method) . . . . . . . . . . . . 1.5 Reconstruction of an Analog Signal from Its Samples . . . . . . 1.5.1 Approximation of the Reconstruction Filter . . . . . . . . 1.6 Discrete Time Processing of Continuous Time Signals . . . . . . 1.7 Continuous Time Processing of Digital Signals . . . . . . . . . . . 1.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.... .... ....

1 2 4

.... .... ....

5 6 7

. . . .

. . . .

. . . .

. . . .

11 25 30 33

. . . . . .

. . . . . .

. . . . . .

. . . . . .

39 45 51 55 61 68

2 Multirate Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Sampling Rate Reduction by an Integer Factor (Downsampling, Compression) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1 Fourier Transform of the Downsampled Signal . . . . . . . . . . 2.1.2 How to Draw the Frequency Response of Downsampled Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.3 Aliasing in Downsampling . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.4 Interpretation of the Downsampling in Terms of the Sampling Period . . . . . . . . . . . . . . . . . . . . . . . . . . . .

71 72 75 78 80 83

vii

viii

Contents

2.1.5 Drawing the Fourier Transform of Downsampled Signal in Case of Aliasing (Practical Method) . . . . . . . . . . . . . . . . 2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Upsampling (Expansion) . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.2 Mathematical Formulization of Upsampling . . . . . . . . . . . . 2.2.3 Frequency Domain Analysis of Upsampling . . . . . . . . . . . . 2.2.4 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.5 Mathematical Analysis of Interpolation . . . . . . . . . . . . . . . . 2.2.6 Approximation of the Ideal Interpolation Filter . . . . . . . . . . 2.2.7 Anti-aliasing Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Practical Implementations of C/D and D/C Converters . . . . . . . . . . 2.3.1 C/D Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Sample and Hold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.3 Quantization and Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.4 D/C Converter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Manipulation of Digital Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.1 Manipulation of Periodic Digital Signals . . . . . . . . . . . . . . . 3.1.2 Shifting of Periodic Digital Signals . . . . . . . . . . . . . . . . . . . 3.1.3 Some Well Known Digital Signals . . . . . . . . . . . . . . . . . . . 3.2 Review of Signal Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Convolution of Periodic Digital Signals . . . . . . . . . . . . . . . . . . . . . 3.3.1 Alternative Method to Compute the Periodic Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Sampling of Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5.1 Aliasing in Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5.2 Matrix Representation of DFT and Inverse DFT . . . . . . . . . 3.5.3 Properties of the Discrete Fourier Transform. . . . . . . . . . . . 3.5.4 Circular Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Practical Calculation of the Linear Convolution . . . . . . . . . . . . . . . 3.6.1 Evaluation of Convolution Using Overlap-Add Method . . . 3.6.2 Overlap-Save Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Computation of the Discrete Fourier Transform . . . . . . . . . . . . . . . 3.7.1 Fast Fourier Transform (FFT) Algorithms . . . . . . . . . . . . . . 3.7.2 Decimation in Time FFT Algorithm . . . . . . . . . . . . . . . . . . 3.7.3 Decimation in Frequency FFT Algorithm . . . . . . . . . . . . . . 3.8 Total Computation Amount of the FFT Algorithm . . . . . . . . . . . . . 3.9 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92 97 97 98 99 103 107 111 126 128 129 130 134 136 139 145 146 149 149 156 158 165 166 170 172 182 184 185 188 198 199 204 207 207 207 217 225 230

Contents

4 Analog and Digital Filter Design . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Review of Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1 Z-Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.2 Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Transformation Between Continuous and Discrete Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Conversion of Transfer Functions of LTI Systems . . . 4.2.2 Forward Difference Transformation Method . . . . . . . . 4.2.3 Bilinear Transformation. . . . . . . . . . . . . . . . . . . . . . . . 4.3 Analogue Filter Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Ideal Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Practical Analog Filter Design . . . . . . . . . . . . . . . . . . 4.3.3 Practical Filter Design Methods . . . . . . . . . . . . . . . . . 4.3.4 Analog Frequency Transformations . . . . . . . . . . . . . . . 4.4 Implementation of Analog Filters . . . . . . . . . . . . . . . . . . . . . . 4.4.1 Low Pass Filter Circuits . . . . . . . . . . . . . . . . . . . . . . . 4.4.2 Analog High-Pass Filter Circuit Design . . . . . . . . . . . 4.4.3 Analog Bandpass Active Filter Circuits . . . . . . . . . . . 4.4.4 Analog Bandstop Active Filter Circuits . . . . . . . . . . . . 4.5 Infinite Impulse Response (IIR) Digital Filter Design (Low Pass) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.1 Generalized Linear Phase Systems . . . . . . . . . . . . . . . 4.6 Finite Impulse Response (FIR) Digital Filter Design . . . . . . . 4.6.1 FIR Filter Design Techniques . . . . . . . . . . . . . . . . . . . 4.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

. . . .

. . . .

. . . .

. . . .

233 233 236 239

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

240 245 246 248 253 254 258 260 272 273 273 279 282 282

. . . . .

. . . . .

. . . . .

. . . . .

283 289 290 291 297

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301

Chapter 1

Sampling of Continuous Time Signals

Signal is a physical phenomenon that carries information. This physical phenomenon is described by mathematical functions, and usually the signal and its mathematical function are used for one another, i.e., synonymous. For instance, when we talk about a sinusoidal signal, we use the sinusoidal function, a mathematical function, to characterize the signal, and the name sinusoidal is used for the signal. Signals are usually depicted in graphs to observe their behavior and analyze them. Sinusoidal signals are the main signals and all the other signals can be considered as being made up of sinusoidal signals with different frequencies and amplitudes. That is to say, any continuous time signal can be written as sum of sinusoidal signals with different frequencies and amplitudes. Rectangular signal, square pulse signal, impulse train signal, triangle signal can be given as examples of continuous time signals. Digital signals are obtained from continuous time signals via sampling operation. Digital signals are represented as mathematical sequences, and the elements of these sequences are nothing but the amplitude values taken from continuous time signals at every multiple of the sampling period. Since in the last several decades a huge improvement is achieved at the development of the digital devices, it has become almost a must especially for electrical engineers to have a good knowledge of digital signals. Digital signals are almost available in every part of our life. Computers, TVs, speakers, mobile phones, house equipment, and most of the other electronic devices process digital signals. In this chapter, we discuss the construction of digital signals via sampling operation, their spectral analyses, the case of aliasing, and reconstruction of a continuous time signal from its samples.

© Springer Nature Singapore Pte Ltd. 2018 O. Gazi, Understanding Digital Signal Processing, Springer Topics in Signal Processing 13, DOI 10.1007/978-981-10-4962-0_1

1

2

1.1

1 Sampling of Continuous Time Signals

Sampling Operation for Continuous Time Signals

Let xc ðtÞ be a continuous time signal. We take samples from the amplitudes of this signal at every multiple of Ts which is called sampling period and form a mathematical sequence. The obtained mathematical sequence is called digital signal. The sampling operation is described by the formula x½n ¼ xc ðnTs Þ

n 2 Z; Ts 2 R

ð1:1Þ

where n is of integer type and Ts is the sampling period. The block diagram of the sampling operation is depicted in Fig. 1.1. Let’s now try to explain the sampling operation on a sinusoidal signal. The graph of the sinusoidal signal with period T is given in Fig. 1.2. Let’s now take some samples from the sine signal in Fig. 1.2, and within this purpose, let’s choose sampling period as Ts ¼ T6 . Samples from signal amplitude are taken at every multiple of Ts , and this operation is illustrated in Fig. 1.3. The sampled amplitude values are placed into an array and expressed as a mathematical sequence. The mathematical sequence obtained from the above sampling operation can be written as  x ½ n ¼

a

b

c d

e

f

g |{z}

h

i

j

k

l



n¼0

which is a digital signal obtained from a continuous time signal. The obtained mathematical sequence can also be displayed graphically as in Fig. 1.4.

Fig. 1.1 Sampling operation of a continuous time signal

Fig. 1.2 Sine signal with period T

1.1 Sampling Operation for Continuous Time Signals

3

Fig. 1.3 Sampling of the sine signal

x [n]

a 6

b

c

5

4

h

d 3

2 e

1

0

g

1

i

2

j 3

4 k

f

5

6

n

l

Fig. 1.4 Digital sine signal

If starting index value, i.e., n ¼ 0, is not indicated in the mathematical sequence, the index of the first element is accepted as n ¼ 0. Graphical illustration is usually employed for easy understanding of the sampling operation and to interpret the meaning of the received signal. Let’s consider the sampling of sine signal again and write a mathematical expression for the digital sine signal. The continuous time sinus signal with period T is written as   2p xc ðtÞ ¼ sin t : T

ð1:2Þ

If the continuous time signal in (1.2) is sampled with sampling period Ts ¼ T6 , we obtain the digital signal x½n whose mathematical expression can be calculated as   p  2p T x½n ¼ xc ðtÞjt¼nTs ! x½n ¼ sin n ! x½n ¼ sin n : T 6 3

ð1:3Þ

By giving negative and positive values to n we obtain the amplitude values of digital sine signal which can be shown as

4

1 Sampling of Continuous Time Signals

2

3

        7 6 2p p 0p 2p 6 7 sin  sin . . .7: x½n ¼ 6. . . sin  sin 4 5 3 3 3 3 |fflffl{zfflffl}

ð1:4Þ

n¼0

Example 1.1 Find the frequency and period of the continuous time signal xc ðtÞ ¼ cosð2ptÞ. Sample the given continuous time signal with sampling period Ts = 1/8 s and obtain the digital signal x½n. Solution 1.1 If xc ðtÞ ¼ cosð2ptÞ is compared to the general form of cosine signal cosð2pftÞ, it is seen that the frequency of xc ðtÞ is f = 1 Hz which can be used to find the period of the signal using T ¼ 1=f leading to T = 1 s. The sampling operation for xc ðtÞ ¼ cosð2ptÞ with sampling period Ts = 1/8 s is done as x½n ¼ xc ðtÞjt¼nTs ! x½n ¼ cosð2ptÞjt¼nTs   1 ! x½n ¼ cos 2pn 8 pn ! x½n ¼ cos : 4

1.1.1

ð1:5Þ

Sampling Frequency

In communication theory; sampling frequency is one of the most important parameters. Sampling frequency is used more than sampling period. Sampling frequency shows the number of samples taken from a continuous time signal per-second. For this reason, it is an indicator of the quality of the continuous-to-digital converters. As sampling frequency increases more samples are taken per-second but this leads to an increase in transmission overhead. As an example, if the sampling frequency is 1000 Hz i.e., 1 kHz, it means that every second, 1000 samples are taken from continuous time signal. Verification Let’s now prove the above claim (the meaning of sampling frequency) for a continuous time periodic signal. Let xc ðtÞ be a continuous time periodic signal, with period T and Ts be the sampling period. In this case, from one period of the signal a total of TTs samples are collected. The continuous time period signal repeats itself T1 times in 1 s. According to this information, in one second, the total number of samples taken from the signal equals to TTs  T1 ! T1s which is nothing but the sampling frequency.

1.1 Sampling Operation for Continuous Time Signals

5

Example 1.2 The continuous time signal xc ðtÞ ¼ cosð2pftÞ where f ¼ 1 kHz is sampled with sampling frequency fs ¼ 16 kHz, and the digital signal x½n ¼ xc ðnTs Þ is obtained. According to the given information, find (a) The number of samples taken from one period of the continuous time signal. (b) The number of samples taken per-second from continuous time signal. Solution 1.2 The number of samples taken per-second from continuous time signal equals the sampling frequency, i.e., fs ¼ 16000 samples are taken per-second. Since 1 the period of the continuous time signal is T ¼ 1 kHz ! T ¼ 1 ms, the number of samples taken from one period of the signal is 16000  1 ms ! 16 samples.

1.1.2

Mathematical Characterization of the Sampling Operation

Impulse Train Impulse train function is one of the most widely used mathematical expression appearing in sampling operation. For this reason, we will first inspect the impulse train function in details. The impulse train function is given as 1 X

sðtÞ ¼

dðt  nTs Þ

ð1:6Þ

n¼1

where Ts is the sampling period. The graph of impulse train function is given in Fig. 1.5. Continuous time periodic signals have Fourier series representation. Impulse train signal (function) also has Fourier series representation which can be written as sðt Þ ¼

1 X

2p

S½k ejkTs t

ð1:7Þ

k¼1

s (t ) 1

5Ts

4Ts

3Ts

Fig. 1.5 Impulse train function

2Ts

Ts

0

Ts

2Ts

3Ts

4Ts

5Ts

t

6

1 Sampling of Continuous Time Signals

where S½k are the Fourier series coefficients which are calculated using S½k ¼

1 1 X 2p sðtÞejkTs t dt: Ts k¼1

ð1:8Þ

Let’s now calculate the Fourier series coefficients of impulse train. Using (1.8) the Fourier series coefficients of the impulse train function can be calculated as S½k ¼

1 1 Z 2p 1 1 dðtÞejkTs t dt ! S½k ¼ e0 ! S½k ¼ Ts 1 Ts Ts

ð1:9Þ

Replacing the calculated coefficients in (1.8) we get the Fourier series representation of the impulse train as sðt Þ ¼

1 2p 1 X ejkTs t Ts k¼1

ð1:10Þ

Using the Fourier series representation of the impulse train function, we can calculate its Fourier transform. For this purpose, we first need to know the Fourier transform of the exponential function. The Fourier transform of the exponential function is given as FT

ejw0 t $ 2pdðw  w0 Þ:

ð1:11Þ

When the expression in (1.11) is used while taking the Fourier transform of (1.10), we obtain the Fourier transform of the impulse train SðwÞ ¼

1.2

1 2p X dðw  kws Þ; Ts k¼1

ws ¼

2p : Ts

ð1:12Þ

Sampling Operation

The first step in sampling operation is to multiply the continuous time signal to be sampled by an impulse train. This multiplication operation for the sampling of sine signal is depicted in Fig. 1.6. When the continuous time signal xc ðtÞ is multiplied by the impulse train sðtÞ; we obtain xs ðtÞ ¼ xc ðtÞ  sðtÞ

ð1:13Þ

in which, if the explicit expression for the impulse train is inserted we get the mathematical expression

1.2 Sampling Operation

7

xs (t )

2Ts 6Ts

5Ts

4Ts

3Ts

Ts

xc (t ) s (t )

4Ts

0

Ts

xs (t )

2Ts

6Ts

5Ts

4Ts

3Ts

Ts

2Ts

5Ts

3Ts

6Ts

t

xc (t ) s (t )

4Ts

0

Ts

2Ts

3Ts

5Ts

6Ts

t

Fig. 1.6 Multiplication of sine signal by an impulse train 1 X

xs ð t Þ ¼ xc ð t Þ

dðt  nTs Þ

ð1:14Þ

n¼1

which can be simplified using the impulse function property R f ðtÞdðt  t0 Þdt ¼ f ðt0 Þ as xs ð t Þ ¼

1 X

xc ðnTs Þdðt  nTs Þ

ð1:15Þ

n¼1

where substituting x½n ¼ xc ðnTs Þ, we obtain xs ð t Þ ¼

1 X

x½ndðt  nTs Þ

ð1:16Þ

n¼1

1.2.1

The Fourier Transform of the Product Signal

We obtained the time domain expression for the product signal xs ðtÞ. Let’s now consider the Fourier transform of the product signal xs ðtÞ. The Fourier transform of xs ðtÞ is computed using

8

1 Sampling of Continuous Time Signals

Z1 Xs ðwÞ ¼

xs ðtÞejwt dt !

1 Z1

Xs ðwÞ ¼ 1

ð1:17Þ

1 X

x½ndðt  nTs Þejwt dt

n¼1

where if the integration and summation expressions are interchanged we get Xs ðwÞ ¼

Z1

1 X

x½n

n¼1

dðt  nTs Þejwt dt

ð1:18Þ

1

on which by using the impulse function properties for the calculation of the integration, Fourier transform of the product signal is obtained as Xs ðwÞ ¼

1 X

x½nejwnTs :

ð1:19Þ

n¼1

The right hand side of the (1.19) contains parameters from time domain. However, there is not only one single expression for the Fourier transform of the product signal. We can find an alternative expression for the Fourier transform of product signal. Let’s now find an alternative expression for the Fourier transform of product signal where both left and right sides only include expressions in frequency domain. Consider the product signal expression again xs ðtÞ ¼ xc ðtÞ  sðtÞ

ð1:20Þ

where the right hand side is the product of two expressions, for this reason, the Fourier transform of xs ðtÞ can be written as Xs ðwÞ ¼

1 Xc ðwÞ  SðwÞ: 2p

ð1:21Þ

where substituting the expression in (1.12) for SðwÞ; we get Xs ðwÞ ¼

1 1 2p X Xc ðwÞ  dðw  kws Þ 2p Ts k¼1

ð1:22Þ

where by using the impulse function property and linearity of the convolution operation we obtain Xs ðwÞ ¼

1 1 X Xc ðw  kws Þ: Ts k¼1

ð1:23Þ

1.2 Sampling Operation

9

We have obtained a second alternative expression for the Fourier transform of product signal. Let’s write both Fourier expressions again Xs ðwÞ ¼

1 X

x½nej2pnTs

Xs ðwÞ ¼

n¼1

1 1 X Xc ðw  kws Þ: Ts k¼1

ð1:24Þ

In these expressions the left hand sides are both Xs ðwÞ. So the right hand sides should also be equal to each other. Equating the right hand sides of the expressions in (1.24), we obtain the equation 1 X

1 1 X Xc ðw  kws Þ: Ts k¼1

x½nejwnTs ¼

n¼1

ð1:25Þ

The Fourier transform of the digital signal x½n is calculated using Xn ðwÞ ¼

1 X

x½nejwn

n¼1

which resembles to the left term in (1.25). We can write the left hand side of (1.25) in terms of Xn ðwÞ as 1 X

x½nejwnTs ¼

n¼1

|fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl}

1 1 X Xc ðw  kws Þ Ts k¼1

ð1:26Þ

Xn ðwTs Þ

which yields Xn ðwTs Þ ¼

1 1 X Xc ðw  kws Þ Ts k¼1

from which Xn ðwÞ can be obtained by replacing w with Xn ðwÞ ¼

w Ts

  1 1 X w Xc  kws : Ts k¼1 Ts

ð1:27Þ and we obtain ð1:28Þ

In the expression (1.28) the left hand side represents the Fourier transform of the digital signal obtained from an analog signal via sampling operation. In other words, it represents the Fourier transform of the mathematical sequence obtained from analog signal via sampling operation. The right hand side consists of shifted and scaled replicas of Xc ðwÞ which is the Fourier transform of analog signal on which sampling operation is performed. Since Xn ðwÞ is the Fourier transform of a digital signal, it is periodic with period 2p. If the digital signal is also periodic in

10

1 Sampling of Continuous Time Signals

time domain, then its Fourier transform is periodic with period 2p consisting of impulses spaced by multiples of 2p. Now, let’s summarize the formulas we have derived up to this point. In Time Domain Continuous time signal Sampling operation Sampling period Impulse train Product signal Product signal Product signal

xc ðtÞ x½n ¼ xc ðnTs Þ Ts P sðt Þ ¼ 1 n¼1 dðt  nTs Þ xs ðtÞ ¼ xP c ðtÞ  sðtÞ xs ð t Þ ¼ 1 x ðnT Þdðt  nTs Þ Pk¼1 c s xs ð t Þ ¼ 1 k¼1 x½ndðt  nTs Þ

In Frequency Domain

R1 xs ðtÞXs ðwÞ ¼ 1 xs ðtÞejwt dt 1 Xc ðwÞ  SðwÞ xs ðtÞXs ðwÞ ¼ 2p 2p ws ¼ Ts P xs ðtÞXs ðwÞ ¼ T1s 1 k¼1 Xc ðw  kws Þ P1 Xn ðwÞ ¼ n¼1 x½nejwn   P w Xn ðwÞ ¼ T1s 1 X  kw c s k¼1 Ts   2 |{z} 3:5 4 5 6 3 2 Exercise: Given the digital signal x½n ¼

Fourier transform of product function Fourier transform of product function Sampling frequency in rad/sec Fourier transform of product function Fourier transform of x½n digital signal Fourier transform of x½n digital signal

n¼0

draw the graphs of P (a) yðtÞ ¼ 1 x½ndðt  nTs Þ where Ts = 1/4 s. Pn¼1 1 (b) gðtÞ ¼ n¼1 x½2ndðt  nTs Þ where Ts = 1/8 s. P (c) hðtÞ ¼ 1 n¼1 x½n=2dðt  nTs Þ where Ts = 1/4 s. Exercise: Calculate the Fourier transforms of x c ð t Þ ¼ dð t Þ þ dð t  1Þ and x½n ¼ d½n þ d½n  1 and draw their magnitude and phase responses.

1.3 How to Draw Fourier Transforms of Product Signal …

1.3

11

How to Draw Fourier Transforms of Product Signal and Digital Signal

The derived mathematical expression for Xs ðwÞ is given as Xs ðwÞ ¼

1 1 X Xc ðw  kws Þ Ts k¼1

ð1:29Þ

which is a periodic function and one period of this function around the origin, assuming no overlapping among shifted replicas, can be written as 1 Xc ðwÞ: Ts

ð1:30Þ

The period of Xs ðwÞ is denoted by ws whose value equals to 2p Ts . Drawing the graph of Xs ðwÞ consists of two steps. In the first step, we draw the graph of T1s Xc ðwÞ around the origin. Then in the next step, the drawn graph around the origin is shifted to the left and right by integer multiples of ws ¼ 2p Ts , i.e., by kws ; k 2 Z, and the shifted replicas together with the one around the origin are all summed. Before studying some problems on the drawing of Xs ðwÞ, let’s inspect some examples to prepare ourselves for the drawing of Xs ðwÞ. Example 1.3 In Fig. 1.7, the graphics of X1 ðwÞ and X2 ðwÞ are given for the interval 0  w  4. Draw the graph of X1 ðwÞ þ X2 ðwÞ for the same interval. Solution 1.3 To draw the graphic of X1 ðwÞ þ X2 ðwÞ, let’s first write the mathematical expressions for each function, then sum these functions to get the mathematical expression for the summed signals. The mathematical expressions for the signals X1 ðwÞ and X2 ðwÞ are given as w w þ 2 X2 ðwÞ ¼ 2 2 If we sum mathematical expressions for the signals X1 ðwÞ and X2 ðwÞ, we get X1 ðwÞ ¼ 

X1 ðwÞ þ X2 ðwÞ ¼ 2:

Fig. 1.7 The graphics of X1 ðwÞ and X2 ðwÞ

X1 ( w)

X2 ( w)

2

0

4

w

12

1 Sampling of Continuous Time Signals

X 1 ( w)

X 1 ( w)

X2 ( w) 2

2

0

X 2 ( w)

4

w

0

w

4

Fig. 1.8 The graphics of X1 ðwÞ þ X2 ðwÞ

Fig. 1.9 The graphics of X1 ðwÞ and X2 ðwÞ functions

X 1 ( w)

X2 ( w)

a

b 0

4

w

The obtained result is graphically shown in Fig. 1.8. Example 1.4 The graphics of X1 ðwÞ and X2 ðwÞ for the interval 0  w  4 are shown in Fig. 1.9. The slopes of the lines in Fig. 1.9 are 1=2 and 1=2. According to the given information, draw the graphic of X1 ðwÞ þ X2 ðwÞ for the same interval. Solution 1.4 To draw the graph of X1 ðwÞ þ X2 ðwÞ we need to find its mathematical expression. For this purpose, let’s first write the mathematical expressions for X1 ðwÞ and X2 ðwÞ using the given information for the interval 0  w  4 as w w þ a X2 ðwÞ ¼ þ b 2 2 When the mathematical expressions for X1 ðwÞ and X2 ðwÞ are summed, we obtain X1 ðwÞ ¼ 

X1 ðwÞ þ X2 ðwÞ ¼ a þ b which is graphically depicted in Fig. 1.10. Example 1.5 The graphics of X1 ðwÞ and X2 ðwÞ functions for the interval 0  w  4 are shown in Fig. 1.11. The slopes of the lines in Fig. 1.11 are 1=2 and 1=2. According to the given information, draw the graphic of X1 ðwÞ þ X2 ðwÞ.

1.3 How to Draw Fourier Transforms of Product Signal …

13

X1 (w) X 2 (w)

X 1 ( w) X 2 ( w)

a b

a b

0

4

w

0

w

4

Fig. 1.10 The graphic of X1 ðwÞ þ X2 ðwÞ Fig. 1.11 The graphic of X1 ðwÞ and X2 ðwÞ

a

X 1 ( w)

X 2 ( w)

b 0

4

w

Solution 1.5 We can follow the same steps as in the previous two examples. The line equations of X1 ðwÞ and X2 ðwÞ can be written as X1 ðwÞ ¼ mw þ a X2 ðwÞ ¼ mw þ b: If we sum the line equations of these two functions, we obtain X1 ðwÞ þ X2 ðwÞ ¼ a þ b: The obtained result is depicted in Fig. 1.12. We will use this result to draw the graphs of the digital signals having the spectral overlapping problem. Example 1.6 xc ðtÞ is a continuous time signal and its Fourier transform is denoted by Xc ðwÞ. The graph of Xc ðwÞ is depicted in Fig. 1.13. As it is seen from the Fourier transform graph, xc ðtÞ is a low-pass signal with bandwidth wN . Let xs ðtÞ ¼ xc ðtÞ  sðtÞ where sðtÞ is the impulse train signal. Draw the Fourier transform of xs ðtÞ assuming that ws [ 2wN , i.e., draw Xs ðwÞ.

14

1 Sampling of Continuous Time Signals

X1 (w) X 2 (w)

a b a

X 1 ( w)

X 2 ( w)

b 0

4

w

0

4

w

Fig. 1.12 The graphic of X1 ðwÞ þ X2 ðwÞ Fig. 1.13 Graph of Xc ðwÞ

Xc (w) A

wN

0

wN

w

Solution 1.6 The Fourier transform of the product signal xs ðtÞ is Xs ðwÞ ¼

1 1 X Xc ðw  kws Þ Ts k¼1

which is a periodic function with period ws ¼ 2p Ts : When the summation expression in Xs ðwÞ is expanded, we get Xs ðwÞ ¼    þ

1 1 1 Xc ðw þ ws Þ þ Xc ðwÞ þ Xc ðw  ws Þ þ    Ts Ts Ts

where the graphs of the terms T1s Xc ðwÞ; T1s Xc ðw þ ws Þ, and T1s Xc ðw  ws Þ are depicted in Fig. 1.14. The other shifted and scaled replicas can be drawn in a similar manner as in Fig. 1.14. When the shifted and scaled replicas are summed, we obtain the graphic of Xs ðwÞ as depicted in Fig. 1.15. Example 1.7 xc ðtÞ is a continuous time signal and its Fourier transform Xc ðwÞ is 1 depicted in Fig. 1.16. xc ðtÞ is sampled by the sampling period Ts ¼ 2000 s. Draw the graph of Xs ðwÞ

1.3 How to Draw Fourier Transforms of Product Signal …

15

1 Xc ( w) Ts A Ts

wN

0

w

wN

1 Xc (w ws ) Ts A Ts

0

ws

wN

ws

ws

wN

w

1 Xc (w ws ) Ts A Ts

ws

wN

ws

ws

wN

Fig. 1.14 The graphics of

1 Ts

w

0

Xc ðwÞ; T1s Xc ðw þ ws Þ and

1 Ts

Xc ðw  ws Þ

Xs (w) A Ts

ws

wN

ws

ws

wN

wN

0

wN

ws

wN

ws

ws

wN

w

Fig. 1.15 The graphic of Xs ðwÞ

Solution 1.7 The sampling frequency in rad/sec is 2p ! ws ¼ 4000p rad=s Ts The shifted Xc ðwÞ signals by multiples of ws are shown in Fig. 1.17. As it is clear from Fig. 1.17, shifted replicas overlap. Summing the overlapped amplitudes, we obtain the signal shown in Fig. 1.18. ws ¼

16

1 Sampling of Continuous Time Signals

X c (w) 1

4000

2000

w

Fig. 1.16 The graphic of Xc ðwÞ

X c (w 4000 )

X c (w)

X c (w 4000 )

1

6000

4000

2000

2000

X c (w 8000 ) X c (w 4000 )

4000

6000

8000

X c (w) X c (w 4000 ) X c (w 8000 ) 1

10000

8000

6000

4000

2000

2000

4000

6000

8000

10000

12000

w

Fig. 1.17 Shifted Xc ðwÞ signals

2 1

10000

8000

6000

4000

2000

2000

4000

6000

8000

10000

12000

w

Fig. 1.18 Summation of the shifted replicas

In the last stage, we divide the amplitudes of the summed signal shown in 1 dividing the amplitudes by Ts equals to multiplying Fig. 1.18 by Ts . Since Ts ¼ 2000 the amplitudes by 2000. After multiplying the amplitudes by 2000 we obtain the graphic of the function Xs ðwÞ as depicted in Fig. 1.19.

1.3 How to Draw Fourier Transforms of Product Signal …

17

X s (w) 4000 2000

10000

8000

6000

4000

2000

2000

4000

6000

8000

10000

w

12000

Fig. 1.19 The graphic of Xs ðwÞ

Xc (w)

Fig. 1.20 The graphic of Xc ðwÞ

1

1000

0

w 1000

Example 1.8 The graphic of Xc ðwÞ is shown in Fig. 1.20. Draw the graphic of Xs ðwÞ ¼ 250

1 X

Xc ðw  k500pÞ

k¼1

Solution 1.8 From the equation Xs ðwÞ ¼ 250

1 X

Xc ðw  k500pÞ

k¼1

it is seen that the sampling frequency in rad/sec is ws ¼ 500p rad=s. Let’s partition the horizontal axis of Xc ðwÞ as in Fig. 1.21 considering the sampling frequency value. Now let’s draw the shifted Xc ðwÞ signals as shown in Fig. 1.22. The graphs of Xc ðwÞ, Xc ðw  ws Þ and Xc ðw þ ws Þ altogether are given in Fig. 1.23. More shifted graphs of Xc ðwÞ are given in Fig. 1.24. If the above graph is carefully inspected, it is seen that a portion of the graph repeats itself along the horizontal axis. The repeated part is indicated by bold lines in Fig. 1.25. Now let’s write the mathematical equations for the line segments, a, b, c, d, e, f, g, h appearing in the repeating pattern in Fig. 1.25 as

18

1 Sampling of Continuous Time Signals

X c (w)

Fig. 1.21 The graphic of Xc ðwÞ

1

1000

500

Fig. 1.22 Shifted graphs of Xc ðwÞ

250

0

250

X c (w)

500

1000

w

X c (w 500 )

1

1000

X c (w

500

250

250

500

w

1000

X c (w)

500 ) 1

1000

500

250

Fig. 1.23 Shifted graphs of Xc ðwÞ X c (w

250

500

X c (w)

500 )

w

1000

X c (w

500 )

1

1000

500

250

250

500

1000

w

1.3 How to Draw Fourier Transforms of Product Signal …

19

1

1000

500

250

250

500

w

1000

Fig. 1.24 Shifted graphs of Xc ðwÞ

Fig. 1.25 Shifted graphs of Xc ðwÞ and repeating pattern

m¼

1 1000p

 250p  w  250p

ya ¼ mw þ 1 ye ¼ mw þ

1 2

yb ¼ mw þ 1 yc ¼ mw þ yf ¼ mw þ

1 2

yg ¼ mw

1 2

yd ¼ mw þ

1 2

yh ¼ mw:

If we sum the equations for the line segments, a, c, e, g and b, d, f, g we get the results ya þ yc þ ye þ yg ¼ 2 yb þ yd þ yf þ yg ¼ 2: and the graph of Xs ðwÞ is drawn as in Fig. 1.26.

Fig. 1.26 Summation result of shifted Xc ðwÞ functions

2

1000

0

w

1000

20

1 Sampling of Continuous Time Signals

Xs (w)

Fig. 1.27 The graphic of Xs ðwÞ

500 w

0

1000

1000

In the last step to get the graph of Xs ðwÞ ¼ 250

1 X

Xc ðw  k500pÞ

k¼1

it is sufficient to multiply the amplitude values of the signal depicted in Fig. 1.26. After amplitude multiplication, we obtain the graph of Xs ðwÞ as depicted in Fig. 1.27. Example 1.9 The graphic of Xc ðwÞ is shown in Fig. 1.28. Draw the graphic of Xs ðwÞ ¼

  1 1 X 2p Xc w  k Ts k¼1 Ts

1 for Ts ¼ 375 s.

Solution 1.9 We can write the sampling frequency in rad=san unit as ws ¼ 2p Ts ! ws ¼ 750prad=s. In the next step, we shift the function Xc ðwÞ to the left and right by kws ; k 2 Z. Some shifted replicas of Xc ðwÞ are displayed in Fig. 1.29. If the graph in Fig. 1.29 is inspected carefully, it can be seen that a define pattern repeats itself along the shape. The repeating pattern is indicated in bold lines in Fig. 1.30. The repeating pattern in Fig. 1.30 is redrawn alone in Fig. 1.31 in details.

X c (w)

Fig. 1.28 The graphic of Xc ðwÞ for Example 1.9

4

1000

500

250

250

500

1000

w

1.3 How to Draw Fourier Transforms of Product Signal …

Xc (w 750 )

1000

750

Xc (w)

4

500

250

250

21

X c (w 750 )

500

750

1000

w

4

1000

500

250

250

500

1000

w

Fig. 1.29 Shifted Xc ðwÞ functions

4

w Fig. 1.30 The repeating pattern

If the graphic in Fig. 1.31 is inspected carefully, it is seen that the line pairs in the upper left and upper right shadowed rectangles overlap each other and their slopes are equal in magnitude but opposite in sign. For this reason, the sum of the line equations for line pairs is a constant number and it equals to 1 + 4 = 5. After summing the overlapping line equations, we get the graphic in Fig. 1.32. If the triangle shape and horizontal line in Fig. 1.32 are summed, we get the graphic in Fig. 1.33. The graphic shown in Fig. 1.33 corresponds to one period of the function Xs ðwÞ around origin. If one period of Xs ðwÞ around origin is shifted to the right and left by multiples of ws ¼ 750p and shifted replicas are all summed together with the graph around origin, we get the graphic of Xs ðwÞ as in Fig. 1.34. Solution 2 In fact, the second solution provided here is more complex than the first solution. However, we find it useful to illustrate the different perspectives for the solution of a problem. The repeating pattern chosen in solution can be interpreted in a different manner. In fact, the interpretation of the repeating patterns depends on the reader’s

22 Fig. 1.31 The repeating pattern drawn in details

Fig. 1.32 The graphic obtained after summing the overlapping lines

Fig. 1.33 The graphic obtained after summing the triangle shape and horizontal line

1 Sampling of Continuous Time Signals

1.3 How to Draw Fourier Transforms of Product Signal …

23

Fig. 1.34 The graphic of Xs ðwÞ: Fig. 1.35 Repeating part in details

perception. The overlapped lines in the repeating pattern are shown inside circles in Fig. 1.35 in a different approach than the one in solution 1. In Fig. 1.35 the sum of the overlapped lines inside circles results in constant numbers, and when the constants are added to the top triangle shape, we obtain one period of Xs ðwÞ around origin. This is illustrated in Fig. 1.36. When the obtained one period around the origin is shifted to the left and right, we obtain Xs ðwÞ function in Fig. 1.37. Exercise: The graphic of Xc ðwÞ function is depicted in Fig. 1.38. Using the given figure draw the graph of

24

1 Sampling of Continuous Time Signals

Fig. 1.36 The sum of the overlapped lines inside circles in repeating pattern

Xs ( w ) 6 5

500

250

250

w

500

Fig. 1.37 Xs ðwÞ graph

X c (w)

Fig. 1.38 Fourier transform of an input signal

2 1

1000

0

500

1000

w

1.3 How to Draw Fourier Transforms of Product Signal … 1 X

Xs ðwÞ ¼ 500

25

Xc ðw  k1000pÞ:

k¼1

Exercise: The Fourier transform of a continuous time signal is given as Xc ðwÞ ¼ pðdðw  500pÞ þ dð þ 500pÞÞ: Using the given Fourier transform draw the graph of Xs ðwÞ ¼ 200

1 X

Xc ðw  k400pÞ:

k¼1

1.3.1

Drawing the Fourier Transform of Digital Signal

Assume that Xn ðwÞ is the Fourier transform of x½n which is obtained from xc ðtÞ via sampling operation, i.e., x½n ¼ xc ðtÞjt¼nTs ! x½n ¼ xc ðnTs Þ and the mathematical expression for Xn ðwÞ is given as   1 1 X w Xn ðwÞ ¼ Xc  kws : Ts k¼1 Ts

ð1:31Þ

To draw the graph of Xn ðwÞ two different methods can be followed. Below, we explain these two methods separately. Method 1: First draw the graph of Xs ðwÞ, i.e., draw the Fourier transform of the product signal xs ðtÞ ¼ xc ðtÞsðtÞ as discussed in the previous section. Once you have the graph of Xs ðwÞ, to get the graph of Xn ðwÞ, multiply the horizontal axis of Xs ðwÞ by sampling period Ts . Method 2: Since Xn ðwÞ is the Fourier transform of the digital signal x½n, it is a periodic signal and  its period equals 2p. To draw the graph of Xn ðwÞ; first draw the

graph of T1s Xc

w Ts

around origin, then shift the drawn signal to the left and right by

multiples   of 2p, and sum the shifted replicas. Note that to draw the graph of 1 w Ts Xc Ts ; we multiply the amplitude values of Xc ðwÞ by 1=Ts and multiply horizontal axis of Xc ðwÞ by Ts , i.e., divide the horizontal axis of Xc ðwÞ by 1=Ts . Let’s now provide some examples to comprehend the subject better. Example 1.10 The Fourier transform of a continuous time signal xc ðtÞ is depicted in Fig. 1.39. Draw Xn ðwÞ, the Fourier transform of x½n ¼ xc ðnTs Þ where Ts is the sampling period. Assume that ws [ 2wN .

26

1 Sampling of Continuous Time Signals

X c ( w) A

0

wN

w

wN

Fig. 1.39 Fourier transform of a low pass input signal

X s (w) A Ts

ws wN

ws

ws

wN

wN

wN

0

ws

wN

ws

ws

wN

w

Fig. 1.40 Graph of Xs ðwÞ

Solution 1.10 Method 1: Let’s first draw the graph of Xs ðwÞ ¼

1 1 X Xc ðw  kws Þ Ts k¼1

which is a periodic function with period ws ¼ 2p Ts : The graph of Xs ðwÞ is shown in Fig. 1.40. In the second step, we multiply the horizontal axis of Xs ðwÞ by Ts to get the graph of Xn ðwÞ. The graph of Xn ðwÞ is shown in Fig. 1.41. X n (w) A Ts

Ts ws

Ts wN

2

Ts ws Ts wN

Fig. 1.41 Graph of Xn ðwÞ

Ts wN

0

Ts wN

Ts ws

Ts wN

2

Ts ws Ts wN

w

1.3 How to Draw Fourier Transforms of Product Signal …

27

1 w Xc ( ) Ts Ts A Ts

0

Ts wN

Fig. 1.42 The graph of

1 Ts

Ts wN

w

  Xc

w Ts

Method 2: In the second method, we first draw the graph of   1 w Xc Ts Ts then we shift the drawn graph to the left and right by multiples of 2p and obtain the  

graph of Xn ðwÞ. To draw the graph of

1 Ts

Xc

w Ts

; we multiply the vertical and

horizontal axes of Xc ðwÞ by and Ts respectively. In Fig. 1.42 the graph of   1 w Ts Xc Ts is depicted.   Let’s denote T1s Xc Tws by Xn1 ðwÞ. To get the graph of Xn ðwÞ, we shift Xn1 ðwÞ to 1 Ts

the left and right by multiples of 2p and sum the shifted replicas. This operation is illustrated in Fig. 1.43. Example 1.11 The continuous time signal xc ðtÞ is given as xc ðtÞ ¼ cosð4000ptÞ: (a) Draw Xc ðwÞ, the Fourier transform of xc ðtÞ. 1 s. Draw (b) Let xs ðtÞ ¼ xc ðtÞsðtÞ where sðtÞ is the impulse train and Ts ¼ 8000 Xs ðwÞ, the Fourier transform of xs ðtÞ. 1 s. Draw Xn ðwÞ, the Fourier transform of (c) Let x½n ¼ xc ðnTs Þ where Ts ¼ 8000 x½n. Solution 1.11 Before computing the Fourier transform of the given cosine signal, let’s review some properties of the exponential signal. The Fourier transform of an exponential signal is given as FT

ejwN t ! 2pdðw  wN Þ

ð1:32Þ

28

1 Sampling of Continuous Time Signals

Fig. 1.43 Xn ðwÞ graph

and sine and cosine signals can be written in terms of the exponential signals as



1 þ jwN t 1 e sinðwN tÞ ¼  ejwN t cosðwN tÞ ¼ e þ jwN t þ ejwN t ð1:33Þ 2j 2 And the Fourier transforms of the sinusoidal signals are given as FT p sinðwN tÞ ! ðdðw  wN Þ  dðw þ wN ÞÞ j FT

ð1:34Þ

cosðwN tÞ ! pðdðw  wN Þ þ dðw þ wN ÞÞ: (a) Since we refreshed some background information we can start to solve our problem. The Fourier transform of xc ðtÞ ¼ cosð4000ptÞ can be calculated as

1.3 How to Draw Fourier Transforms of Product Signal …

29

FT

cosð4000ptÞ ! pðdðw  4000pÞ þ dðw þ 4000pÞÞ and its graph is depicted as in Fig.P1.44. (b) Since xs ðtÞ ¼ xc ðtÞsðtÞ, and sðtÞ ¼ 1 k¼1 dðt  kTs Þ, where Ts is the sampling period, Fourier transform of xs ðtÞ is Xs ðwÞ ¼

1 1 X Xc ðw  kws Þ Ts k¼1

2p where ws ¼ 2p Ts ! 1=8000 ¼ 16000p. Using the Fourier transform expression Xc ðwÞ found in the previous part, Xs ðwÞ can be calculated as

Xs ðwÞ ¼ 8000

1 X

Xc ðw  k16000pÞ !

k¼1 1 X

Xs ðwÞ ¼ 8000p

ðdðw  4000p  k16000pÞ þ dðw þ 4000p  k16000pÞÞ

k¼1

and the graph of Xs ðwÞ is displayed in Fig. 1.45.

Xc (w)

4000

4000

w

Fig. 1.44 Fourier transform of xc ðtÞ ¼ cosð4000ptÞ

Xs (w) 8000

16000

4000

4000

Fig. 1.45 Fourier transform of the product signal xs ðtÞ

16000

w

30

1 Sampling of Continuous Time Signals

Fig. 1.46 Fourier transform of Xn ðwÞ

(c) To get the graph of Xn ðwÞ; it is sufficient to multiply the horizontal axis of Xs ðwÞ by Ts . Thus, the graph of Xn ðwÞ is obtained as in Fig. 1.46.

1.4

Aliasing (Spectral Overlapping)

Let the Fourier transform of a continuous time signal be as given as in Fig. 1.47. Using the Fourier transform in Fig. 1.47, let’s draw the graph of Xs ðwÞ ¼

1 1 X Xc ðw  kws Þ Ts k¼1

ð1:35Þ

as in Fig. 1.48. It is clear from Fig. 1.48 that the condition for the shifted graphs not to overlap can be written as ws  w1 [ w2 ! ws [ w1 þ w2

ð1:36Þ

and if w1 \w2 then no aliasing condition in (1.36) can also be written as ws [ 2w2 . If ws \w1 þ w2 , then the shifted graphs overlap and this condition is named as aliasing (overlapping). The case of aliasing is depicted in Fig. 1.49.

Fig. 1.47 The Fourier transform of a low pass signal

1.4 Aliasing (Spectral Overlapping)

31

Fig. 1.48 The graph of Xs ðwÞ

Fig. 1.49 Aliasing case

For many signals the Fourier transform is symmetric with respect to the vertical axis, i.e., w1 ¼ w2 . And for the symmetric case, let w1 ¼ w2 ¼ wN and the condition for no aliasing in this case can be stated as ws [ 2wN

ð1:37Þ

where the unit of the frequencies is rad/sec. If we write the explicit expressions for the frequencies in (1.37), we get 2p 2p [2 Ts TN

ð1:38Þ

and the condition for no aliasing can be written as fs [ 2fN . This means that for no aliasing, the sampling frequency in unit of Hertz should be greater than twice of the highest frequency available in the signal.

32

1 Sampling of Continuous Time Signals

Note: If Xc ðwÞ is a complex function, to see the overlapping case graphically we first draw the graph of jXc ðwÞj, and then the graph of Xs ðwÞ ¼

1 1 X jXc ðw  kws Þj Ts k¼1

ð1:39Þ

is drawn. Example 1.12 The Fourier transform of continuous time signal is shown in Fig. 1.50. Draw the Fourier transform of the product signal xs ðtÞ ¼ xc ðtÞsðtÞ and decide on the aliasing case. P Solution 1.12 The graph of Xs ðwÞ ¼ T1s 1 k¼1 Xc ðw  kws Þ is depicted in Fig. 1.51. It is clear from Fig. 1.51 that for no overlapping, we should have ws  wN [ wN

ð1:40Þ

ws [ 2wN

ð1:41Þ

leading to

Fig. 1.50 Graph of Xc ðwÞ

X s (w) A Ts

ws

wN

ws

ws

Fig. 1.51 Graph of Xs ðwÞ

wN

wN

0

wN

ws

wN

ws

ws

wN

w

1.4 Aliasing (Spectral Overlapping)

33

and no aliasing condition in (1.41) can also be expressed as ws [ 2wN ! 2pfs [ 2  2pfN ! fs [ 2fN :

1.4.1

ð1:42Þ

The Meaning of the Aliasing (Overlapping)

Sampling frequency implies the number of samples taken per-second from a continuous time signal. The collected samples are either transmitted, stored, or processed, and the analog signal can be reconstructed from the digital samples. If sampling frequency is not high enough, the analog signal cannot be reconstructed due to insufficient number of received samples or it can only be partially reconstructed. In frequency domain, the effect of insufficient number of samples is seen as aliasing or spectral overlapping. Example 1.13 The continuous time signal xc ðtÞ ¼ cosð20ptÞ þ sinð40ptÞ is to be sampled. Choose a sampling frequency such that no aliasing occurs for the generated digital signal in frequency domain. Solution 1.13 Let’s first calculate the Fourier transform of the continuous time signal. For this purpose, the Fourier transforms of sinusoidal signals are reminded as FT

cosðw0 tÞ $ pðdðw  w0 Þ þ dðw þ w0 ÞÞ FT p sinðw0 tÞ $ ðdðw  w0 Þ  dðw þ w0 ÞÞ j where substituting w0 ¼ 2pf0 ; w ¼ 2pf , we get the alternative form for the Fourier transform of the sinusoidal signals as 1 ðdðf  f0 Þ þ dðf þ f0 ÞÞ 2 FT 1 sinð2pf0 tÞ $ ðdðf  f0 Þ  dðf þ f0 ÞÞ 2j While obtaining the alternative forms, we made use of the property FT

cosð2pf0 tÞ $

dð2pðf  f0 ÞÞ ¼

1 dðf  f0 Þ: 2p

ð1:43Þ

Using the Fourier transform formulas for the sinusoidal signals, we can calculate the Fourier transform of the continuous time signal given in the example and plot its graph as in Fig. 1.52.

34

1 Sampling of Continuous Time Signals

Fig. 1.52 Fourier transform of the composite signal xc ðtÞ

X c (w)

40

20

0

20

40

w

Fig. 1.53 Graph of jXc ðwÞj

The Fourier transform of the summed sinusoids given in Fig. 1.52 seems to be complex to judge although not impossible. For easiness of the illustration, let’s take the absolute value of the Fourier transforms and depict them as in Fig. 1.53. As it is seen from Fig. 1.53 that the highest frequency available in the continuous time signal xc ðtÞ is 40p rad/s or 20 Hz and the lowest positive frequency is 0. The analog signal is a low pass signal. The sampling frequency preventing aliasing should satisfy ws [ 2  40p or in terms of unit of Hz, fs [ 40 Hz.

1.4 Aliasing (Spectral Overlapping)

35

Method 2: Comparing the given sinusoidal functions to cosð2pf1 tÞ and sinð2pf2 tÞ expressions, we find the frequencies of the sinusoidal signals as f1 ¼ 10 Hz and f2 ¼ 20 Hz, and decide on the sampling frequency as fs [ 2  20 Hz ! fs [ 40 Hz Example 1.14 If x½n ¼ xc ðnTs Þ then the Fourier transform of x½n is written as Xn ðwÞ ¼

  1 1 X w Xc  kws Ts k¼1 Ts

ð1:44Þ

where Xc ðwÞ is the Fourier transform of continuous time signal xc ðtÞ. The Fourier transform of the digital signal x½n can also be calculated using the Fourier transform formula directly, i.e., 1 X

Xn ðwÞ ¼

x½nejwn

ð1:45Þ

n¼1

Derive (1.44) starting from the right hand side of (1.45). Solution 1.14 Before starting to the derivation, let’s remember the Fourier and inverse Fourier transforms of continuous time signal Z1 X c ðw Þ ¼

xc ðtÞe

jwt

1 xc ð t Þ ¼ 2p

dt

1

Z1 Xc ðwÞejwt dw: 1

If the time parameter ‘t’ is replaced by ‘nTs’ in inverse Fourier transform expression, we get 1 xc ðnTs Þ ¼ 2p

Z1 Xc ðwÞejwnTs dw:

ð1:46Þ

1

For the digital signal x½n, we have the Fourier transform expression Xn ðwÞ ¼

1 X

x½nejwn

ð1:47Þ

n¼1

in which if we substitute x½n ¼ Xc ðnTs Þ, we get Xn ðwÞ ¼

1 X

xc ðnTs Þejwn :

n¼1

In (1.48) if xc ðnTs Þ is replaced by (1.46), we get

ð1:48Þ

36

1 Sampling of Continuous Time Signals

Xn ðwÞ ¼

1 X 1 1 Z Xc ðkÞejknTs dkejwn 2p 1 n¼1

ð1:49Þ

1 1 1 X Z X c ð kÞ ejðwkTs Þn dk 2p n¼1 1

ð1:50Þ

which can be re-arranged as Xn ðwÞ ¼

and exchanging the places of summation and integration operators, we obtain Xn ðwÞ ¼

1 X 1 1 Z Xc ðkÞ ejðwkTs Þn dk 2p 1 n¼1

ð1:51Þ

on which we can use the property 1 X

ejðwkTs Þn ¼ 2p

1 X

dðw  kTs  k2pÞ

ð1:52Þ

   w 2p dðw  kTs  k2pÞ ¼ d Ts kk Ts Ts   1 w 2p kk ¼ d Ts Ts Ts

ð1:53Þ

  1 2p X w 2p d kk Ts k¼1 Ts Ts

ð1:54Þ

n¼1

1 X

k¼1

ejðwkTs Þn ¼

n¼1

leading to the expression   1 1 1 Z 2p X w 2p Xn ðwÞ ¼ X c ð kÞ d kk dk 2p 1 Ts k¼1 Ts Ts

ð1:55Þ

where upon exchanging summation and integration operators, we get   1 1 X 1 1 Z w 2p Xc ðkÞd kk dk Xn ðwÞ ¼ Ts k¼1 2p 1 Ts Ts

ð1:56Þ

in which the integration expression can be simplified using the impulse function property 1 Z 1

Xc ðkÞdðk0  kÞdk ¼ Xc ðk0 Þ

ð1:57Þ

1.4 Aliasing (Spectral Overlapping)

37

Fig. 1.54 xc ðtÞ graph for Example 1.15

xc (t ) 1

T

t

T

as follows     1 1 Z w 2p 1 w 2p Xc Xc ðkÞd kk k dk ¼ 2p 1 Ts Ts 2p Ts Ts

ð1:58Þ

Finally, when (1.58) is used in (1.56), we get the desired final expression as   1 1 X w 2p Xn ðwÞ ¼ Xc k : Ts k¼1 Ts Ts

ð1:59Þ

Exercise: The inverse Fourier transform for digital signals is given as x ½ n ¼

1 Z Xn ðwÞejwn dw: 2p 2p

ð1:60Þ

Starting from the right hand side of (1.60) and replacing Xn ðwÞ in (1.60) by (1.59) obtain the left hand side of (1.60). Example 1.15 The time domain signal given in Fig. 1.54 is to be sampled. Determine the sampling frequency such that the digital signal contains sufficient information about analog signal and analog signal can be reconstructed from the digital samples. Solution 1.15 To determine the sampling frequency, we need to know the largest and smallest positive frequencies available in the signal spectrum. For this purpose, we calculate the Fourier transform of the continuous time signal and determine the largest and smallest positive frequencies available in the signal spectrum. The Fourier of the continuous time signal is computed as

38

1 Sampling of Continuous Time Signals

X c (w) 2T

w

0 5

4

3

2

T

T

T

T

T

2

3

4

5

T

T

T

T

T

Fig. 1.55 Fourier transform of xc ðtÞ in Fig. 1.54

X c ðw Þ ¼

1 Z

xc ðtÞejwt dt

1

¼

ZT

1ejwt dt

T jwT

 ejwT jw 2 sinðwT Þ ¼ w The graph of the Fourier transform is depicted in Fig. 1.55. Since ¼

e

Xc ð0Þ ¼

0 0

the value of the Fourier transform at origin can be computed using the L’Hôpital’s rule. If we take the derivatives of numerator and denominator of Xc ðwÞ w.r.t w and evaluate it for w ¼ 0, we obtain    dXc ðwÞ 2TcosðwT Þ dXc ðwÞ ¼  ! dw  ¼ 2T dw w¼0 1 w¼0 w¼0 which is nothing but the value of Xc ðwÞ at origin, i.e., Xc ð0Þ. As it is seen from Fig. 1.55, the largest positive frequency in the signal spectrum goes to infinity and the smallest non-negative frequency is 0. We need to choose infinity as sampling frequency and this is not a feasible value for practical implementations. However, as it is seen from the Fourier transform graph, the amplitude of the signal spectrum decreases sharply when frequency is beyond Tp . So, we can assume that the spectrum amplitude is negligible beyond a frequency value. We can

1.4 Aliasing (Spectral Overlapping)

39

choose the largest frequency as wN ¼ 4p T , and according to the chosen frequency, we can write the lower bound for sampling frequency as ws [ 2wN ! ws [ 2 ws [

4p T

8p 2p 8p T ! ! Ts \ [ T Ts T 4 fs [

4 T

Let’s assume that the sampling period is chosen as Ts ¼ T8 . This means that we take 2T T ¼ 16 samples from rectangle signal per second. And these 16 samples are 8

sufficient for reconstruction of the rectangle signal.

1.4.2

Drawing the Frequency Response of Digital Signal in Case of Aliasing (Practical Method)

In sampling operation if the sampling frequency is chosen as fs \2wN where wN is the bandwidth of the low pass analog signal, then aliasing occurs in Fourier transform of the digital signal x½n, i.e., in graph of Xn ðwÞ. The relations between digital signal and continuous time signal in time and frequency domains are as x½n ¼ xc ðtÞjt¼nTs ! x½n ¼ xc ðnTs Þ   1 1 X w Xc  kws : Xn ðwÞ ¼ Ts k¼1 Ts Let the Fourier transform of the continuous time signal to be sampled be as in Fig. 1.56.   If fs \2wN , then the graph of T1s Xc Tws happens to be as in Fig. 1.57. 1 Ts

IfFig.  1.57 is inspected carefully it is seen that when fs \2wN , the function w Xc Ts takes values outside the interval ðp; pÞ on horizontal axis. In Fig. 1.58,

the shadowed triangles denoted by ‘A’ and ‘B’ show the intervals outside ðp; pÞ   where the function T1s Xc Tws has nonzero value.

40

1 Sampling of Continuous Time Signals

Xc (w)

Fig. 1.56 Fourier transform of a continuous time signal

A

wN

Fig. 1.57 Graph of

1 Ts

0

wN

w

  Xc

w Ts

Fig. 1.58  The graph of 1 Ts

Xc

w Ts

If the shadowed triangles ‘A’ and ‘B’ in Fig. 1.58 are shifted to the right and left by 2p, we obtain the graphic in Fig. 1.59. If the overlapping lines in Fig. 1.59 are summed, we obtain the graphic shown in bold lines in Fig. 1.60. As it is clear from Fig. 1.60, due to the overlapping regions the original signal is spectrum is destroyed.

1.4 Aliasing (Spectral Overlapping)

41

Fig. 1.59 Shifting of the shadowed triangles

Fig. 1.60 Summation of the overlapping lines

The amount of this destruction depends   on the widths of the shadowed triangles.

In other words, as the function T1s Xc

w Ts

extends outside the interval ðp; pÞ more,

the amount of distortion on the original signal due to overlapping increases. The graph obtained after summing the overlapping lines is depicted alone in Fig. 1.61. Let’s now, step by step, describe drawing the graph of Xn ðwÞ in case of aliasing in an easy and practical manner.   Step 1: First we draw the graph of T1s Xc Tws . For this purpose, we divide the horizontal axis of the graph of Xc ðwÞ by 1=Ts i.e., we multiply the horizontal axis by Ts , and multiply the amplitude values by 1=Ts . Step 2: If the sampling frequency is chosen as fs \2wN , then aliasing occurs in the Fourier  transform of x½n, i.e., aliasing occurs in Xn ðwÞ. And in this case, the graph  of

1 Ts

Xc

w Ts

extends beyond the interval ðp; pÞ. The portion of the graph

extending to the left of p is denoted by ‘A’, and the potion extending to the right of p is denoted by ‘B’.

42

1 Sampling of Continuous Time Signals

Fig. 1.61 The resulting graph after summing the overlapping lines

Step 3: The portion of the graph denoted by ‘A’ in Step 2 is shifted to the right by 2p, and the portion denoted by ‘B’ is shifted to the left by 2p. The overlapping lines are summed and one period of Xn ðwÞ around origin is obtained. Let’s denote this one period by Xn1 ðwÞ. Step 4: In the last step, one period of Xn ðwÞ around origin denoted by Xn1 ðwÞ is shifted to the left and right by multiples of 2p and all the shifted replicas are summed to get Xn ðwÞ, this is mathematically stated as 1 X

Xn ðwÞ ¼

Xn1 ðw  k2pÞ:

k¼1

Example 1.16 The Fourier transform of continuous time signal xc ðtÞ is shown in Fig. 1.62. This signal is sampled and digital signal x½n ¼ xc ðtÞjt¼nTs ! x½n ¼ xc ðnTs Þ, Ts ¼ 1=64 is obtained. Draw the graph of the Fourier transform digital signal, i.e., draw the graph of Xn ðwÞ. Solution 1.16 Step 1: First we draw the graph of

1 Ts

  Xc

w Ts

, for this purpose, we multiply the

horizontal axis of Xc ðwÞ in Fig. 1.62 by Ts ¼ 1=64 and multiply the vertical axis of Xc ðwÞ in Fig. 1.62 by 1=Ts ¼ 64. The resulting graph is shown in Fig. 1.63. Fig. 1.62 Fourier transform of a low pass input signal

1.4 Aliasing (Spectral Overlapping)

43

Fig. 1.63  The graph of 1 Ts

Xc

w Ts

Fig. 1.64  The graph of 1 Ts

Xc

w Ts

Fig. 1.65 The portions of graph outside ðp; pÞ interval are labelled by ‘A’ and ‘B’

The graph in Fig. 1.63 is drawn more in details as in Fig. 1.64 where we see that the graph   extends to the outside of the (p; p) interval. And in fact, the parts of the 1 Ts

Xc

w Ts

extending beyond (p; p) cause the spectral overlapping problem due to

the 2p periodicity of Xn ðwÞ: Step 2: We shadow the portion of the graphs outside the ðp; pÞ interval and denote them by the letters ‘A’ and ‘B’, we obtain the graph in Fig. 1.65. If the shadowed portions labelled by ‘A’ and ‘B’ are shifted to the right and to the left by 2p, we obtain the graph in Fig. 1.66. In Fig. 1.66, we can write the equations of the overlapping lines for the interval 256 192 ðp; 3p=4Þ as 128 3p w þ 64 and  5p w  5 , and when these two equations are 128 summed, we obtain  15p w þ 128 5 . In a similar manner, if we write the equations of the overlapping lines for the interval ðp=2; pÞ and sum them, we obtain

44

1 Sampling of Continuous Time Signals

Fig. 1.66 Shadowed portions are shifted to the right and to the left by 2p

Fig. 1.67 One period of Xn ðwÞ around origin

Fig. 1.68 Fourier transform of a continuous time signal

128  15p w þ 128 3 . After summing the overlapping line equations, we can draw one period of Xn ðwÞ around origin as in Fig. 1.67. Step 3: In the last step, we shift one period of Xn ðwÞ around origin to the left and right by multiples of 2p and summing all the non-overlapping shifted replicas, we obtain the graph of Xn ðwÞ.

Exercise: The Fourier transform of a continuous time signal xc ðtÞ is depicted in Fig. 1.68. This signal is sampled with sampling period Ts ¼ 1=32 and digital signal x½n is obtained. Draw the Fourier transform of x½n. Exercise: The Fourier transform of a continuous time signal xc ðtÞ is depicted in Fig. 1.69. This signal is sampled with sampling period Ts ¼ 1=32 and digital signal x½n is obtained. Draw the Fourier transform of x½n.

1.5 Reconstruction of an Analog Signal from Its Samples

45

Fig. 1.69 Fourier transform of a continuous time signal

1.5

Reconstruction of an Analog Signal from Its Samples

To obtain a digital signal x½n from an analog signal xc ðtÞ via sampling operation, we first multiply the analog signal by an impulse train sðtÞ and obtain the product signal xs ðtÞ ¼ xc ðtÞsðtÞ. Then we collect the amplitude values of impulses from xs ðtÞ and form the digital sequence x½n. Now we wonder the reverse operation, i.e., assume that we have the digital sequence x½n, then how can we construct the analog signal xc ðtÞ? To achieve this, we will just follow the reverse operations. That is, we will first obtain xs ðtÞ from x½n, then from xs ðtÞ we will extract xc ðtÞ. Let’s study the reconstruction operation in time domain as shown in Fig. 1.70. As it is depicted in Fig. 1.70, we can write mathematical expression for the product signal xs ðtÞ in terms of the elements of digital signal x½n but we have no way to write an expression for xc ðtÞ using xs ðtÞ. Hence, we cannot solve the reconstruction problem in time domain. Let’s inspect the reconstruction operation in frequency domain then. Assume that xc ðtÞ is a low pass signal and its Fourier transform is as given in Fig. 1.71. Considering the Fourier transform in Fig. 1.71, we can draw the Fourier transform of the product signal xs ðtÞ as in Fig. 1.72. The Fourier transform of xs ðtÞ is a periodic signal with period ws and it’s one period around origin equals to T1s Xc ðwÞ in case of no aliasing. It is clear from Fig. 1.72 that for no aliasing, we should have ws [ 2wN !

2p [ 2wN Ts

Fig. 1.70 Reconstruction operation in time domain

ð1:61Þ

46

1 Sampling of Continuous Time Signals

Fig. 1.71 Fourier transform of xc ðtÞ

Fig. 1.72 Fourier transform of xs ðtÞ

Fig. 1.73 Multiplication of Xs ðwÞ by rectangle function Hr ðwÞ

2p p [ 2wN ! wN \ : Ts Ts

ð1:62Þ

Now consider the reconstruction operation in frequency domain. We had problem in converting xs ðtÞ to xc ðtÞ in time domain. However, it is clear from Fig. 1.72 that it is easy to get the Fourier transform of xc ðtÞ, i.e., Xc ðwÞ from the Fourier transform of xs ðtÞ, i.e., Xs ðwÞ. To get Xc ðwÞ from Xs ðwÞ, it is sufficient to multiply Xs ðwÞ by a rectangle function centered around the origin. This operation is depicted in Fig. 1.73 where rectangle function is denoted by Hr ðwÞ which is nothing but the transfer function of a low pass analog filter.

1.5 Reconstruction of an Analog Signal from Its Samples

47

Fig. 1.74 Fourier transform of the reconstruction filter

The Fourier transform of the low pass analog filter is depicted in Fig. 1.74 alone. In fact, the filter under consideration is an ideal lowpass filter, and it is used just to illustrate the reconstruction operation. In practice, such ideal filters are not available, and practical non-ideal filters are employed for reconstruction operations. The time domain expression of the analog filter with the frequency response depicted in Fig. 1.74 can be calculated using the inverse Fourier transform formula as follows: hr ð t Þ ¼

1 1 Z Hr ðwÞejwt dw 2p 1 p

¼

1 TZs Ts ejwt dw 2p  p Ts

Ts jwt Tps e p ¼ Ts 2p Ts jTp t p

e s  ejTs t ¼ j2pt

where using the property sinðhÞ ¼ 2j1 ejh  ejh ; we obtain   hr ð t Þ ¼

sin

pt Ts pt Ts

:

ð1:63Þ

Since sin cð xÞ ¼

sinðpxÞ px

ð1:64Þ

the mathematical expression in (1.63) can be written in terms of sin cðÞ function as

48

1 Sampling of Continuous Time Signals

Fig. 1.75 Reconstruction filter impulse response

  t hr ðtÞ ¼ sin c : Ts

ð1:65Þ

The graph of the reconstruction filter hr ðtÞ is depicted in Fig. 1.75 where it is clear that the reconstruction filter takes 0 value at every multiple of Ts . As we explained before the Fourier transform of the continuous time signal can be written as the multiplication of Xs ðwÞ and Hr ðwÞ i.e., Xc ðwÞ ¼ Xs ðwÞHr ðwÞ:

ð1:66Þ

Since multiplication in frequency domain equals to convolution in time domain, (1.66) can be also be expressed as xc ðtÞ ¼ xs ðtÞ  hr ðtÞ

ð1:67Þ

where substituting 1 X

x½ndðt  nTs Þ

n¼1

for xs ðtÞ, we obtain xc ð t Þ ¼ ¼

1 X n¼1 1 X

x½ndðt  nTs Þ  hr ðtÞ ð1:68Þ x½nhr ðt  nTs Þ

n¼1

which is nothing but the reconstruction expression of the analog signal xc ðtÞ.

1.5 Reconstruction of an Analog Signal from Its Samples

49

Note: f ðtÞ  dðt  t0 Þ ¼ f ðt  t0 Þ Using (1.63) in (1.68) the reconstructed analog signal from its samples can be written as

xc ð t Þ ¼

1 X

x½n

  sÞ sin p ðtnT Ts

n¼1

pðtnTs Þ Ts

ð1:69Þ

or in terms of sin cðÞ function, it is written as xc ð t Þ ¼

1 X n¼1

 x½n sin c

t  nTs Ts

 ð1:70Þ

Example 1.17 The continuous time signal xc ðtÞ ¼ sinð2ptÞ is sampled by sampling period Ts ¼ 14 s. (a) Write the digital sequence x½n obtained after sampling operation. (b) Assume that x½n is transmitted and available at the receiver. Reconstruct the analog signal at the receiver side from its samples, i.e., using x½n reconstruct the analog signal xc ðtÞ. Solution 1.17 (a) The frequency of the sinusoidal signal xc ðtÞ ¼ sinð2ptÞ is 1 Hz, and its period is 1 s. Sampling period is Ts ¼ 14 s. Every multiple of Ts , we take a sample from the sinusoidal signal. The graph of the sinusoidal signal and the samples taken from its one period are indicated in Fig. 1.76. Since sampling frequency is fs ¼ 4 Hz, we take 4 samples per-second from the signal. The samples taken from one period of the sinusoidal signal can be written as ½ 0 1 0 1 . Since the sine signal is defined from 1 to 1. The obtained digital signal is a periodic signal and in this digital signal, the repeating pattern happens to be ½ 0 1 0 1 . The digital signal obtained from the sampling operation can be written as

Fig. 1.76 Sampling of sine signal

50

1 Sampling of Continuous Time Signals

2 6... 0 x ½ n ¼ 4

1

0

3

n¼0

z}|{ 1 0 1 0 |fflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflffl}

1

0

1

0

1

...7 5

Repeating pattern

ð1:71Þ (b) At the receiver, the analog signal can be reconstructed from its samples using xc ð t Þ ¼

1 X

x½nhr ðt  nTs Þ

ð1:72Þ

n¼1

where Ts ¼ 14 and   hr ð t Þ ¼

sin

pt Ts pt Ts

:

ð1:73Þ

Using the x½n in (1.72), the reconstructed signal can be written as xc ðtÞ ¼    þ hr ðt þ 3Ts Þ  hr ðt þ Ts Þ þ hr ðt  Ts Þ  hr ðt  3Ts Þ þ   

ð1:74Þ

The graph of hr ðtÞ in (1.73) is depicted in Fig. 1.77 where it is clear that the amplitude of the main lobe of hr ðtÞ equals to 1, and the function equals to 0 when t is a multiple of Ts . The shifted copies of hr ðtÞ and their summation is illustrated in Fig. 1.78.

Fig. 1.77 Reconstruction filter impulse response

1.5 Reconstruction of an Analog Signal from Its Samples

51

Fig. 1.78 Summing the shifted sin cðÞ functions to reconstruct the analog signal

If we only pay attention to the main lobes in Fig. 1.78, we see that the reconstruction signal resembles to the sine signal. Overlapping tails improve the accuracy of the reconstructed signal.

1.5.1

Approximation of the Reconstruction Filter

The reconstruction filter hr ðtÞ ¼ sin c

  t Ts

is depicted in Fig. 1.79 where it is seen

that the filter has a large main lobe and small side lobes, and as the time values hr (t ) 1

t

0 5Ts 4Ts 3Ts 2Ts Ts

Fig. 1.79 Reconstruction filter impulse response

Ts 2Ts 3Ts 4Ts 5Ts

52

1 Sampling of Continuous Time Signals

hr (t )

har (t )

1

1 Approximation

Ts

Ts

Ts

t

0

0

Ts

t

Fig. 1.80 Approximation of the reconstruction filter

increase, the amplitudes of the side lobes decrease. To construct a simplified model for the reconstruction filter, we can approximate the lobes by isosceles triangles. In Fig. 1.80 the main lobe of the reconstruction filter is approximated by an isosceles triangle and the side lobes are all omitted. This type of approximation can also be called as linear approximation. For the triangle in Fig. 1.80, we can write line equations for the left and right edges. For the left edge, the line equation is t þ 1; Ts

Ts  t\0;

for the right edge, the line equation is 

t þ 1; Ts

0  t  Ts

and combining these two line equations into a single expression, we can write the linearly approximated filter expression as har ¼

 Tjtsj þ 1 0

0  j t j  Ts : otherwise

Example 1.18 The continuous time signal xc ðtÞ ¼ sinð2ptÞ is sampled by sampling period Ts ¼ 14.

1.5 Reconstruction of an Analog Signal from Its Samples

53

(a) Write the digital sequence x½n obtained after sampling operation. (b) Assume that x½n is transmitted and available at the receiver. Reconstruct the analog signal at the receiver side from its samples using approximated reconstruction filter. Solution 1.18 (a) We solved this problem before and found the digital signal as 2 6... 0 x½n ¼ 4

1

0

1

3

n¼0

z}|{ 0 1 0 1 |fflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflffl}

0 1

0

1 . . . 7 5:

Repeating pattern

ð1:75Þ (b) At the receiver side, the analog signal can be reconstructed from its samples using xc ð t Þ ¼

1 X

x½nhar ðt  nTs Þ

ð1:76Þ

n¼1

where Ts ¼ 14 and har ðtÞ is the approximated reconstruction filter. Using the x½n found in the previous part, and expanding (1.76), the reconstructed signal can be written as

Fig. 1.81 Reconstruction of analog signal using approximated filter

54

1 Sampling of Continuous Time Signals

xc ðtÞ ¼    þ har ðt þ 3Ts Þ  har ðt þ Ts Þ þ har ðt  Ts Þ  har ðt  3Ts Þ þ    ð1:77Þ The shifted copies of har ðtÞ in (1.77) and their summation is illustrated in Fig. 1.81. As it is seen from Fig. 1.81, the reconstructed signal resembles to the sine signal. Now we ask the question: How can we obtain a better reconstructed sine signal? Answer Either we can use a better approximated filter or take more samples from one period of the signal, i.e., increase the sampling frequency which means, decrease the sampling period. To get a better approximated filter, we can represent the side-lobes by the small triangles. A better approximation of the reconstruction filter is illustrated in Fig. 1.82 where it is seen that two side lobes are approximated by triangles. Although improved linear approximation improves the accuracy of the reconstructed signal, the sharp discontinuities of the linear approximated filter makes the realization of the filter difficult. Reconstruction operation can be illustrated using block diagrams as in Fig. 1.83.

Fig. 1.82 Better approximation of the reconstruction filter

Fig. 1.83 Reconstruction operation using block diagram

1.5 Reconstruction of an Analog Signal from Its Samples

In Fig. 1.83, if hr ðtÞ ¼ sin c

55

  t Ts ; then perfect reconstruction occurs, i.e.,

xr ðtÞ ¼ xc ðtÞ.

1.6

Discrete Time Processing of Continuous Time Signals

Currently most of the electronic devices are produced using digital technology. For this reason, analog signals are usually converted to digital signals and processed by digital electronic systems. These electronic units can be digital filters, equalizers, amplifiers, etc. In Fig. 1.84, the general system for digital processing of analog system is depicted. The system in Fig. 1.84 can be inspected both in time and frequency domains assuming that discrete time system is linear and time invariant. Let’s first write the relations among signals in time, and then in frequency domain. Time Domain Relations: x½n ¼ xc ðnTs1 Þy½n ¼ x½n  h½n yr ðtÞ ¼

1 X

y½nhr ðt  nTs2 Þ

ð1:78Þ

n¼1

If perfect reconstruction filter is to be employed, then 

 t hr ðtÞ ¼ sin c : Ts2

ð1:79Þ

Frequency Domain Relations:   1 1 X w Xn ðwÞ ¼ Xc  kws1 Ts1 k¼1 Ts1

ð1:80Þ

where ws1 ¼

2p ; Ts1

Yn ðwÞ ¼ Xn ðwÞHn ðwÞ:

ð1:81Þ

To write the frequency domain relation between y½n and yr ðtÞ, let’s remember the two-stage reconstruction process illustrated as follows

Fig. 1.84 Digital processing of a continuous time signal

56

1 Sampling of Continuous Time Signals

We have Ys ðwÞ ¼

1 1 X Yc ðw  kws2 Þ Ts2 k¼1

 Yn ðwÞ ¼ Ys

By

w Ts2

Yn ðwÞ ¼

  1 1 X w Yc  kws2 Ts2 k¼1 Ts2

ð1:82Þ

 ! Yr ðwÞ ¼ Hr ðwÞYs ðwÞ ! Yr ðwÞ ¼ Hr ðwÞYn ðTs2 wÞ: ð1:83Þ

combining

Xn ðwÞ ¼ T1s

1 P

 Xc

k¼1

w Ts

  kws ; Yn ðwÞ ¼ Xn ðwÞHn ðwÞ

and

Yr ðwÞ ¼ Hr ðwÞYn ðTs2 wÞ, we get the relation between Yr ðwÞ and Xc ðwÞ as   1 1 X Ts2 Yr ðwÞ ¼ Hr ðwÞHn ðTs2 wÞ Xc w  kws1 Ts1 k¼1 Ts1

ð1:84Þ

If Ts1 ¼ Ts2 ¼ Ts , then (1.84) reduces to Yr ðwÞ ¼ Note: Hr ðwÞ ¼

Ts 0

Ts Hn ðTs wÞXc ðwÞ; 0;

if  Tps  w  otherwise

 Tps  w  otherwise

p Ts

:

ð1:85Þ

p Ts

Example 1.19 In Fig. 1.85, the graphs of Xc ðwÞ and Xn ðwÞ are depicted. In addition, x½n ¼ xc ðtÞjt¼nTs . By comparing the graphs of Xc ðwÞ and Xn ðwÞ, write Xc ðwÞ in terms of Xn ðwÞ.

Fig. 1.85 Graphs for Example 1.19

1.6 Discrete Time Processing of Continuous Time Signals

57

Fig. 1.86 One period of Xn ðwÞ around origin

Solution 1.19 First let’s write the expression for one period of Xn ðwÞ around origin as Xn ðwÞ

 p  w\p

ð1:86Þ

which is graphically shown as in Fig. 1.86. If we divide the horizontal axis of Xn ðwÞ by Ts , we get Xn ðTs wÞ



p p  w\ Ts Ts

ð1:87Þ

which is graphically depicted in Fig. 1.87. If we multiply the amplitudes by Ts , we obtain Ts X n ð T s w Þ



p p  w\ : Ts Ts

ð1:88Þ

which is graphically depicted in Fig. 1.88. Figure 1.88 is nothing but the graph of Xc ðwÞ. As a result, we can conclude that if x½n ¼ xc ðtÞjt¼nTs , then we can express Fourier transform of xc ðtÞ i.e., Xc ðwÞ in terms of Fourier transform of x½n i.e., Xn ðwÞ as Xc ðwÞ ¼ Ts Xn ðTs wÞ

Fig. 1.87 One period of Xn ðTs wÞ around origin



p p  w\ : Ts Ts

ð1:89Þ

58

1 Sampling of Continuous Time Signals

Fig. 1.88 One period of Ts Xn ðTs wÞ around origin

Fig. 1.89 Continuous to digital converter

Example 1.20 For the continuous to digital converter given in Fig. 1.89, assume that the sampling frequency is high enough so that there is no aliasing in frequency domain. Xn ðwÞ is the Fourier transform of x½n, and Xc ðwÞ is the Fourier transform of xc ðtÞ. Write one period of Xn ðwÞ in terms of Xc ðwÞ. Solution 1.20 Since Xn ðwÞ is the Fourier transform of a digital signal, Xn ðwÞ is periodic and its period equals 2p, the relation between Xn ðwÞ and Xc ðwÞ is given as   1 1 X w 2p Xn ðwÞ ¼ Xc k Ts k¼1 Ts Ts

ð1:90Þ

which is written explicitly as       1 w 2p 1 w 1 w 2p Xc þ þ Xc þ Xc  þ    ð1:91Þ Ts Ts Ts Ts Ts Ts Ts Ts |fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl} n¼1 n¼0 n¼1   1 w In (1.91), let Yc ðwÞ ¼ Ts Xc Ts , then it is obvious that Yc ðw  2pÞ ¼   1 w 2p X  c Ts Ts Ts . The explicit expression of Xn ðwÞ can be written as Xn ðwÞ ¼    þ

Xn ðwÞ ¼    þ Yc ðw þ 2pÞ þ Yc ðwÞ þ Yc ðw  2pÞ þ   

ð1:92Þ

From (1.92), it is obvious   that one period of Xn ðwÞ is Yc ðwÞ, that is to say, one 1 period of Xn ðwÞ is Ts Xc Tws and this can mathematically be written as

1.6 Discrete Time Processing of Continuous Time Signals

  1 w X n ðw Þ ¼ X c Ts Ts

59

 p  w\p

ð1:93Þ

p p  w\ : Ts Ts

ð1:94Þ

which can also be written as Ts Xn ðwTs Þ ¼ Xc ðwÞ



Example 1.21 For the digital to continuous converter given in Fig. 1.90, let Yn ðwÞ be the Fourier transform of y½n. Because Yn ðwÞ is the Fourier transform of a digital signal, it is periodic and its period equals 2p. Let Ynop ðwÞ be the one period of Yn ðwÞ around origin. That is Ynop ðwÞ ¼ Yn ðwÞ  p  w\p. Write the Fourier transform of yr ðtÞ, i.e., Yr ðwÞ in terms of Ynop ðwÞ. Solution 1.21 Digital to continuous conversion operation is reminded in Fig. 1.91. As a result, we can write the relation between one period of Yn ðwÞ and Yr ðwÞ as Yr ðwÞ ¼ Ts Ynop ðTs wÞ

ð1:95Þ

The expression in (1.95) can also be written as Yr ðwÞ ¼ Ts Yn ðTs wÞ

Fig. 1.90 Digital to continuous converter

Fig. 1.91 Digital to continuous conversion



p p w : Ts Ts

ð1:96Þ

60

1 Sampling of Continuous Time Signals

Example 1.22 If x½n ¼ Ts yc ðtÞjt¼nTs , write one period of the Fourier transform of x½n in terms of Fourier transform of yc ðtÞ. Assume that there is no aliasing. Solution 1.22 Using the expression below   1 Ts X w 2p Xn ðwÞ ¼ Yc k Ts Ts Ts k¼1

ð1:97Þ

the relation in one period can be written as   w ! Yc ðwÞ ¼ Xn ðTs wÞ: Xn ðwÞ ¼ Yc Ts

ð1:98Þ

Example 1.23 In Fig. 1.92 two signal processing systems are depicted. If both systems produce the same output yr ðtÞ for the same input signal xc ðtÞ, find the relation between the impulse responses of continuous time and discrete time systems. Solution 1.23 For the first system, the frequency domain relation between system input and output is Yr ðwÞ ¼ Hc ðwÞXc ðwÞ ð1:99Þ Considering only one period (op) of the Fourier transforms of the digital signals around origin, the relations between input and output of each unit can be written as C/D:   1 w Xnop ðwÞ ¼ Xc Ts Ts

ð1:100Þ

Ynop ðwÞ ¼ Hn ðwÞXnop ðwÞ

ð1:101Þ

Disc.Time System:

Fig. 1.92 Signal processing systems for Example 1.23

1.6 Discrete Time Processing of Continuous Time Signals

61

D/C: Yr ðwÞ ¼ Ts Ynop ðTs wÞ

ð1:102Þ

If we combine the expressions (1.100–1.102), we get Yr ðwÞ ¼ Hnop ðTs wÞXc ðwÞ:

ð1:103Þ

If we equate the right hand sides of the Eqs. (1.99) and (1.103), we get   w Hc ðwÞ ¼ Hnop ðTs wÞ ! Hnop ðwÞ ¼ Hc Ts

ð1:104Þ

from which we can write the time domain relation for h½n and hc ðtÞ as h½n ¼ Ts hc ðtÞjt¼nTs :

1.7

ð1:105Þ

Continuous Time Processing of Digital Signals

Digital signals can be processed by continuous time systems. For this purpose, the digital signal is first converted to continuous time signal then processed by a continuous time system whose output is back converted to a digital signal. The overall procedure is depicted in Fig. 1.93. For the system in Fig. 1.93, time and frequency domain relations between block inputs and outputs are as follows: Time domain relations are   1 X t  nTs xc ð t Þ ¼ x½n sin c ð1:106Þ yc ðtÞ ¼ xc ðtÞ  hc ðtÞ Ts n¼1 y½n ¼ yc ðtÞjt¼nTs :

ð1:107Þ

Frequency domains relations are X c ðw Þ ¼

Ts Xn ðTs wÞ 0

Fig. 1.93 Continuous time processing of digital signals

if  Tps  w  otherwise

p Ts

Yc ðwÞ ¼ Xc ðwÞHc ðwÞ

ð1:108Þ

62

1 Sampling of Continuous Time Signals

Fig. 1.94 Signal processing units for Example 1.24

Fig. 1.95 Signal processing units for Example 1.24

  1 1 X w 2p Yn ðwÞ ¼ Yc k : Ts k¼1 Ts Ts

ð1:109Þ

Example 1.24 The signal processing units given in Figs. 1.94 and 1.95 have the same outputs for the same given inputs. Find the relation between the impulse responses of discrete and continuous time systems. Solution 1.24 For the first system, the relation between input and output is Yn ðwÞ ¼ Hn ðwÞXn ðwÞ: ð1:110Þ Here Yn ðwÞ is periodic with period 2p and one period of it can be written as either Yn ðwÞ ¼ Hn ðwÞXn ðwÞ

 p  w\p

ð1:111Þ

or Ynop ðwÞ ¼ Hn ðwÞXnop ðwÞ

ð1:112Þ

For the second system, the relations between block inputs and outputs are given as X c ðw Þ ¼

Ts Xn ðTs wÞ 0

if  Tps  w  otherwise

Yn ðwÞ ¼

p Ts

Yc ðwÞ ¼ Xc ðwÞHc ðwÞ

  1 1 X w 2p Yc k : Ts k¼1 Ts Ts

ð1:113Þ ð1:114Þ

If we combine the expressions in (1.113), we get Y c ðw Þ ¼

Hc ðwÞTs Xn ðTs wÞ if  Tps  w  0 otherwise

and substituting (1.115) into (1.114), we obtain

p Ts

ð1:115Þ

1.7 Continuous Time Processing of Digital Signals

Yn ðwÞ ¼

1 X k¼1

 Hc

 w 2p k Xn ðw  k2pÞ: Ts Ts

63

ð1:116Þ

One period of Yn ðwÞ is Ynop ðwÞ ¼ Hc

  w Xnop ðwÞ Ts

ð1:117Þ

If we equate the right hand sides of (1.112) and (1.117) Hn ðwÞXnop ðwÞ ¼ Hc

    w w Xnop ðwÞ ! Hn ðwÞ ¼ Hc Ts Ts

ð1:118Þ

which is can be expressed in time domain as h½n ¼ Ts hc ðtÞjt¼nTs :

ð1:119Þ

Example 1.25 Sample continuous time signal in Fig. 1.96, and reconstruct the continuous time signal from its samples. Use triangle approximated reconstruction filter during reconstruction process. Solution 1.25 The Fourier transform graph of a rectangle signal of length 2T around origin is repeated in Fig. 1.97. For our example; T ¼ 2, let’s choose the approximate bandwidth of the rectangle pulse as wN ¼ 2p=T ! wN ¼ 2p=2 ! wN ¼ p. We can choose the sampling frequency according to ws [ 2wN

ws [ 2p

ð1:120Þ

as 2pfs [ 2p ! fs [ 1 ! fs ¼ 2

ð1:121Þ

which means that the sampling period is Ts ¼ 12. The sampling operation of the rectangle pulse is depicted in Fig. 1.98. The digital sequence obtained after sampling of the rectangular signal is

Fig. 1.96 Continuous time signal for Example 1.25

64

1 Sampling of Continuous Time Signals

Fig. 1.97 Fourier transform of a rectangle signal

Fig. 1.98 Sampling of the rectangular signal

Fig. 1.99 Linear approximation of the reconstruction filter

 x½n ¼

1

1

1

1

1 |{z} n¼0

1

1

1

 :

ð1:122Þ

The construction of the approximated reconstruction filter is repeated in Fig. 1.99.

1.7 Continuous Time Processing of Digital Signals

65

Fig. 1.100 Linear approximation of the reconstruction filter for Ts ¼ 12

Note that our sampling period is Ts ¼ 12, then the approximated reconstruction filter becomes as in Fig. 1.100. Now we can start the reconstruction operation, the reconstruction expression is given as xr ð t Þ ¼

1 X

x½nhar ðt  nTs Þ

ð1:123Þ

n¼1

where Ts ¼ 12 s, and using our digital signal x½n and expanding the summation in (1.123), we obtain         4 3 2 1 þ har t þ þ har t þ þ har t þ xr ðtÞ ¼ har t þ 2 2 2 2       1 2 3 þ har ðtÞ þ har t  þ har t  þ har t  : 2 2 2 The shifted filters appearing in (1.124) is depicted in Fig. 1.101.

Fig. 1.101 Shifted triangle reconstruction filters

ð1:124Þ

66

1 Sampling of Continuous Time Signals

Fig. 1.102 Sum of the shifted reconstruction filters Fig. 1.103 Reconstructed 1 signal for Ts ¼ 16

If the shifted graphs given in Fig. 1.101 are summed, we get the resulting graph shown in bold lines in Fig. 1.102. In Fig. 1.103 the reconstructed signal is depicted alone. As it is seen from Fig. 1.103, the reconstructed signal resembles to the rectangle signal given in the exercise. However, at the left and right sides we have some problems. To increase the accuracy of the reconstructed signal, we should either take more samples from the continuous time signal or increase the accuracy of the reconstruction filter. Let’s take more samples. For this reason, we can increase the sampling frequency, meaning, decrease the sampling period. Accordingly, we can choose Ts ¼ 1=16, which means that we take ð2  ð2ÞÞ  16 ¼ 64 samples from the given continuous time signal. The triangular approximated reconstruction filter for this new sampling period is shown in Fig. 1.104. As it is seen from Fig. 1.104, the edges of the triangle have larger slopes in magnitude. It is not difficult to see from Fig. 1.104 that as the sampling frequency goes to infinity, the reconstruction filter converges to impulse function. Applying the same steps for the new sampling period, we find the reconstructed signal as in Fig. 1.105.

1.7 Continuous Time Processing of Digital Signals

67

Fig. 1.104 Linear approximation of the reconstruction filter for Ts ¼ 12

Fig. 1.105 Reconstructed 1 signal for Ts ¼ 16

Fig. 1.106 Signal graph for Example 1.26

As it is seen from Fig. 1.105, we have a better reconstructed signal. Left and right edges of the reconstructed signal have larger slopes. Note: If unit is not provided for sampling period or for signal axis we accept it as “second” by default. Example 1.26 Is the signal given in Fig. 1.106 a digital signal? Solution 1.26 Time axis of a digital signal consist of only integers. For the given signal, real values appear along time axis. Hence, the signal is not a digital signal but it is discrete amplitude continuous time signal. In fact the signal consists of shifted impulses dðt  t0 Þ which is a continuous function.

68

1.8

1 Sampling of Continuous Time Signals

Problems

(1) For the sampling periods Ts ¼ 1 s and Ts ¼ 1:5 s, draw the graph of sðt Þ ¼

1 X

dðt  nTs Þ:

n¼1

(2) The signal depicted in Fig. 1.107 is sampled. (a) For the sampling period Ts ¼ 1 s, first draw the graph of impulse train function sðtÞ, then draw the graph of the product signal xs ðtÞ ¼ xc ðtÞsðtÞ. Find the digital signal x½n and draw its graph. (b) For the sampling period Ts ¼ 0:5 s repeat part (a) (3) For the impulse train function sðt Þ ¼

1 X

dðt  nTs Þ

n¼1

find (a) Fourier series coefficients. (b) Fourier series representation. (c) Fourier transform. (4) If xs ðtÞ ¼ xc ðtÞsðtÞ where sðtÞ is the impulse train and xc ðtÞ is a continuous time signal, derive the Fourier transform expression of xs ðtÞ in terms of the Fourier transform of xc ðtÞ. (5) If x½n ¼ xc ðnTs Þ, then derive the expression for the Fourier transform of x½n in terms of the Fourier transform of xc ðtÞ. (6) Write mathematical equation for the lines depicted in Fig. 1.108, and then find the sum of these line equations. (7) xc ðtÞ ¼ cosð8ptÞ is sampled and x½n ¼ xc ðnTs Þ digital signal is obtained. According to this information, answer the following. (a) If the sampling period is Ts ¼ 14 s, write the mathematical sequence consisting of the samples taken from the interval 0  t  1. 1 (b) Repeat the previous part for the sampling period Ts ¼ 16 s. 1 1 (c) Which sampling period is preferred Ts ¼ 4 s or Ts ¼ 16 s? Fig. 1.107 Continuous time signal

1.8 Problems

69

Fig. 1.108 Two lines for Question-6

b

a

0

t

Fig. 1.109 Fourier transform of a continuous time signal

Fig. 1.110 Fourier transform of a continuous time signal

1 (8) The continuous time signal xc ðtÞ is sampled with sampling period Ts ¼ 5000 s and the digital signal x½n ¼ xc ðnTs Þ is obtained. The Fourier transform of the continuous time signal is depicted in Fig. 1.109. Draw the Fourier transform of the digital signal x½n. (9) If Ts ¼ 18 s and x½n ¼ ½ 2 3 5 1 2 3 1:5 4:3 2:5 2:5 2 , then draw the graph of

xs ð t Þ ¼

1 X

x½ndðn  Ts Þ:

n¼1

(10) The Fourier transform of a continuous time signal is depicted in Fig. 1.110. Using inverse Fourier transform formula, calculate the time domain expression of this signal.

70

1 Sampling of Continuous Time Signals

Fig. 1.111 Continuous time signal graph for Question-13

Fig. 1.112 Fourier transform of a continuous time signal

(11) Let xs ðtÞ be the product of xc ðtÞ and the impulse train function sðtÞ. Using the product signal expression, write the mathematical expression for the reconstructed signal which is evaluated as xr ðtÞ ¼ xs ðtÞ  hr ðtÞ. (12) For the sampling period Ts ¼ 18 s, draw the linearly approximated reconstruction filter graph. (13) The graph of the continuous time signal xc ðtÞ is displayed in Fig. 1.111. The signal xc ðtÞ is sampled with sampling periods Ts ¼ 1 s, Ts ¼ 14 s and Ts ¼ 18 s. Find the digital signal x½n for each sampling period. (14) A continuous time signal is sampled with sampling period Ts ¼ 18 s and the digital signal x½n ¼ ½ 1 0:7 0 0:7 1 0:7 0 0:7  is obtained. Using the approximated triangle reconstruction filter, rebuild the continuous time signal. (15) The Fourier transform of a continuous time signal xc ðtÞ is depicted in Fig. 1.112. 1 s and The continuous time signal is sampled with sampling period Ts ¼ 3000 the digital signal x½n ¼ xc ðtÞjt¼nTs is obtained. Draw the Fourier transform of x½n. (16) The continuous time signal xc ðtÞ ¼ cosð2p  100  tÞ þ cosð2p  400  tÞ is sampled with sampling frequency fs . How should fs be chosen such that no aliasing occurs in the spectrum of digital signal. (17) A continuous time signal is sampled with sampling frequency fs ¼ 1000 Hz. How many samples per second are taken from continuous time signal?

Chapter 2

Multirate Signal Processing

Digital signals are obtained from continuous time signals via sampling operation. Continuous time signals can be considered as digital signals having infinite number of samples. Sampling is nothing but selecting some of these samples and forming a mathematical sequence called digital signal. And these digital signals can be in periodic or non-periodic forms. The number of samples taken from a continuous time signal per-second is determined by sampling frequency. As the sampling frequency increases, the number of samples taken from a continuous time signal per-second increases, as well. As the technology improves, new and better electronic devices are being produced. This also brings the compatibility problem between old and new devices. One such problem is the speed issue of the devices. Consider a communication device transmitting digital samples taken from a continuous time signal at a high speed. This means high sampling frequency, as well. If the speed of the receiver device is not as high as the speed of the transmitter device, then the receiver device cannot accommodate the samples taken from the transmitter. This results in communication error. Hence, we should be able to change the sampling frequency according to our needs. We should be able to increase or decrease the sampling frequency without changing the hardware. We can do this using additional hardware components at the output of the sampling devices. One way of decreasing the sampling frequency is the elimination of some of the samples of the digital signal. This is also called sampling of digital signals, or decimation of digital signals, or compression of digital signals. On the other hand, after digital transmission, at the receiver side before digital to analog conversion operation, we can increase the number of samples. This is called upsampling, or increasing sampling rate, or increasing sampling frequency. If we have more samples for a continuous time signal, when it is reconstructed from its samples, we obtain a better continuous time signal. In this chapter, we will learn how to manipulate digital signals, which means, changing their sampling rates, reconstruction of a long digital sequence from a short version of it, de-multiplexing and multiplexing of digital signals via hardware units etc. © Springer Nature Singapore Pte Ltd. 2018 O. Gazi, Understanding Digital Signal Processing, Springer Topics in Signal Processing 13, DOI 10.1007/978-981-10-4962-0_2

71

72

2 Multirate Signal Processing

2.1

Sampling Rate Reduction by an Integer Factor (Downsampling, Compression)

To represent a continuous time by digital sequences, we take samples from the continuous time signal according a sampling frequency and form a mathematical sequence. If the mathematical sequence contains too many samples, we can omit some of these samples and keep the rest of the samples for transmission, storage, processing etc. Let’s give another example from real life. Assume that you want to send 500 students to a university in a foreign country. The selected students represent your university and from each department 10 students were selected. Later on you think that the travel cost of 500 students is too much and decide on reducing the number of selected students. A continuous time signal can be considered as a digital signal containing infinite number of samples for any time interval. Sampling of analog signals is nothing but selecting a finite number of samples from the infinite sample sets of the analog signals for the given time interval. The downsampling operation can be considered as the sampling of digital signals. In this case a digital signal containing a number of samples for a given time interval is considered and for the given interval, some of the samples of the digital signal are selected and a new digital signal is formed. This operation is called downsampling. During the downsampling some of the samples of a digital signal are selected and the remaining samples are omitted. The downsampling operation is illustrated in Fig. 2.1 where x½n is the signal to be downsampled and y½n is the signal obtained after downsampling x½n, i.e., after omitting sampled from x½n, and M is the downsampling factor. Given x½n to find the compressed signal, i.e., downsampled signal, y½n, we divide the time axis of x½n by M and keep only integer division results and omit all non-integer division results. Let’s illustrate this operation by an example. Example 2.1 A digital signal expressed as a mathematical sequence is given as x½n ¼ ½3:3

 2:5

1:2

4:5

5:5

 2:3

5:0 |{z}

6:2

3:4

2:3

 4:4

3:2

2:0

n¼0

Find the downsampled y½n ¼ x½3n. Fig. 2.1 Downsampling operation

x [n]

M

y[ n ] = x[ Mn ]

2.1 Sampling Rate Reduction by an Integer Factor …

73

Solution 2.1 Let’s write the time index values of the signal, x½n explicitly follows x½n ¼ ½|{z} 3:3

2:5 |ffl{zffl}

1:2 |ffl{zffl}

4:5 |{z}

5:5 |{z}

2:3 |ffl{zffl}

5:0 |{z}

6:2 |{z}

3:4 |{z}

n¼6

n¼5

n¼4

n¼3

n¼2

n¼1

n¼0

n¼1

n¼2

2:3 |{z}

4:4 |ffl{zffl}

3:2 |{z}

n¼3

n¼4

n¼5

2:0 : |{z} n¼6

In the second step, we divide the time axis of x½n by 3, this is illustrated in ½|{z} 3:3

2:5 |ffl{zffl}

1:2 |ffl{zffl}

4:5 |{z}

5:5 |{z}

2:3 |ffl{zffl}

5:0 |{z}

6:2 |{z}

3:4 |{z}

2:3 |{z}

n¼63

n¼53

n¼43

n¼33

n¼23

n¼13

n¼03

n¼13

n¼23

n¼33

4:4 |ffl{zffl}

3:2 |{z}

n¼43

n¼53

2:0 : |{z} n¼63

where divisions’ yielding integer results are shown in bold numbers and these divisions are given alone as follows ½|{z} 3:3

4:5 |{z}

5:0 |{z}

2:3 |{z}

2:0  |{z}

n¼63

n¼33

n¼03

n¼33

n¼63

and when the divisions are done, we obtain the downsampled signal as y½n ¼ ½|{z} 3:3

4:5 |{z}

5:0 |{z}

2:3 |{z}

2:0  |{z}

n¼2

n¼1

n¼0

n¼1

n¼1

As it is seen from the previous example, downsampling a digital signal by M means that from every M samples of the digital signal only one of them is selected and the rest of them are eliminated. As an example, if y½n ¼ x½6n, then from every 6 samples of x½n only one of them is kept and the other 5 samples are omitted. Now we ask the question, if sampling frequency is fs and downsampling factor is M, after downsampling operation how many samples per-second are available at the downsampler output? The answer is given in the block diagram in Fig. 2.2. Where d:e is the upper-floor operation. If fs is a multiple of M, the diagram in Fig. 2.2 reduces to the one in Fig. 2.3. Example 2.2 Interpret the block diagram given in Fig. 2.4. Solution 2.2 At the input of the block, we receive 300 samples per-second which are obtained from an analog signal via sampling operation. At the output of the downsampler only 1 of every 3 samples is kept and the other 2 samples are omitted.

Fig. 2.2 Sampling frequency at the downsampler output

fs

M

⎡ fs ⎤ ⎢M ⎥ ⎢ ⎥

74

2 Multirate Signal Processing

fs M

M

fs

Fig. 2.3 Sampling frequency at the downsampler output when fs is a multiple of M

f s = 300

3

fs

f ds =

3

→ f ds = 100

Fig. 2.4 Downsampler for Example 2.2

That means at the output of the downsampler, 100 samples every per-second are released. Example 2.3 Find the Fourier series representation of p½n ¼

1 X

d½n  rM:

ð2:1Þ

r¼1

Solution 2.3 The given signal is a periodic signal with period M. Its Fourier series coefficients are computed as M þ1

Mþ1

2 2 1 X 1 X 1 2p 2p P½k ¼ p½nej M kn ! P½k  ¼ d½nej M kn ! P½k ¼ : M M M M1 M1

n¼

n¼

2

ð2:2Þ

2

Using the Fourier series coefficients in (2.2), the Fourier series representation of (2.1) can be written as p½ n ¼

X

2p

P½k ej M kn ! p½n ¼

k;M

The mathematical expression p½n ¼  p½n ¼

1 0

P1 r¼1

1 X j2pkn eM : M k;M

ð2:3Þ

d½n  rM can also be written as

if n ¼ 0; M; 2M; . . . otherwise:

ð2:4Þ

And equating the right hand sides of (2.3) and (2.4) to each other, we get the equality 1 X 1M 2p ej M kn ¼ M k¼0



1 0

n ¼ 0; M; 2M; . . . otherwise:

ð2:5Þ

2.1 Sampling Rate Reduction by an Integer Factor …

75

For the expression in (2.5), if we change the sign of n appearing on both sides of the equation, we obtain an alternative expression for (2.5) as 1 X 1M 2p ej M kn ¼ M k¼0

2.1.1



n ¼ 0; M; 2M; . . . otherwise:

1 0

ð2:6Þ

Fourier Transform of the Downsampled Signal

Let’s find the Fourier transform of the compressed signal y½n ¼ x½Mn. The Fourier transform of y½n can be calculated using Yn ðwÞ ¼

1 X

x½Mnejwn

ð2:7Þ

n¼1

where defining r , Mn, we obtain X

Yn ðwÞ ¼

r

x½r ejwM

ð2:8Þ

r¼0;M;2M

which can be written after parameter changes as X

Yn ðwÞ ¼

n

x½nejwM

ð2:9Þ

n¼0;M;2M

The frontiers of the sum symbol in (2.9) can be changed to 1 and 1 if (2.1) is used in (2.9) as 1 X

Yn ðwÞ ¼

x½n

n¼1

where replacing

P1 r¼1

1 X

n

d½n  rMejwM

r¼1

d½n  rM by its Fourier series representation, we get

Yn ðwÞ ¼

1 X n¼1

x½n

1 X j2pkn jw n e M e M M k;M

ð2:10Þ

which can be rearranged as Yn ðwÞ ¼

1 1X X w þ k2p x½nej M n M k;M n¼1 |fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl} ¼Xn ðw þMk2pÞ

ð2:11Þ

76

2 Multirate Signal Processing

The expression in (2.11) can be reduced to Yn ðwÞ ¼

  1X w þ k2p Xn : M k;M M

ð2:12Þ

In (2.10), if (2.5) was used, then we would obtain Yn ðwÞ ¼

  1X w  k2p : Xn M k;M M

ð2:13Þ

Hence, considering (2.12) and (2.13), we can write the Fourier transform of y½n ¼ x½Mn as Yn ðwÞ ¼

  1 X 1M w  k2p : Xn M k¼0 M

ð2:14Þ

Example 2.4 If y½n ¼ x½Mn the relation between Fourier transforms of x½n and y½n is given as   1 X 1M w  k2p Xn : Yn ðwÞ ¼ M k¼0 M Using the inverse Fourier transform expression for y½n, i.e., Z2p

1 y½n ¼ 2p

Yn ðwÞejwn dw w¼0

show that y½n ¼ x½Mn. Solution 2.4 The inverse Fourier transform is given as 1 y ½ n ¼ 2p

Z2p Yn ðwÞejwn dw 0

where inserting   1 X 1M w  k2p Yn ðwÞ ¼ Xn M k¼0 M

ð2:15Þ

2.1 Sampling Rate Reduction by an Integer Factor …

77

we get 1 X 1 M y ½ n ¼ 2pM k¼0



Z2p Xn

 w þ k2p jwn e dw M

ð2:16Þ

0

In (2.16), if we let k ¼ w þMk2p, then dw ¼ Mdk, and changing the frontiers of the integral (2.16) reduces to ðk þ 1Þ2p M

Z

1 X 1 M y ½ n ¼ 2p k¼0

Xn ðkÞejMkn dk:

ð2:17Þ

k2p M

If (2.17) is expanded for all k values, we obtain ZM

2p

y ½ n ¼

1 2p

ZM

4p

Xn ðkÞejMkn dk þ

1 2p

0

Xn ðkÞejMkn dk þ    2p M

ð2:18Þ

1 2p Z Xn ðkÞejMkn dk þ 2p M1 M

where using the property expression

Rb a

2p

ðÞ þ

Rc b

ðÞ ¼

1 y ½ n ¼ 2p

Rc a

ðÞ and changing k with w, we get the

Z2p Xn ðwÞejMwn dw: 0

When (2.19) is compared to 1 x ½ n ¼ 2p

Z2p Xn ðwÞejwn dw 0

it is seen that y½n ¼ x½Mn.

ð2:19Þ

78

2 Multirate Signal Processing

2.1.2

How to Draw the Frequency Response of Downsampled Signal

To draw the graph of   1 X 1M w  k2p Yn ðwÞ ¼ Xn M k¼0 M students usually expand the summation as Yn ðwÞ ¼

    1 w 1 w  2p 1 w  4p Xn þ Xn þ Xn þ  M M M M M M

ð2:20Þ

and try to draw each shifted graph and sum the shifted graphs. However, this approach is too time consuming and error-prone. Instead of this approach, we will suggest a simpler method to draw the graph of Yn ðwÞ as explained in the following lines. Since Yn ðwÞ is the Fourier transform of the digital signal y½n, then Yn ðwÞ is a periodic signal and its period equals to 2p. To draw the graph of Yn ðwÞ, we can follow the following steps. Step 1: First one period of Xn ðwÞ around origin is drawn. For this purpose, the frequency interval is chosen as p\w  p. Step2: Considering one period of Xn ðwÞ around origin, we draw one period of 1 w 1 w . To draw (in one period) the graph of , we multiply the horizontal X X M n M M n M axis of Xn ðwÞ by M, and multiply the vertical axis of Xn ðwÞ by M1 : Step 3: In Step 3, we shift the resulting graph in Step 2 to the left and right by multiples of 2p and sum the shifted replicas. Let’s now give an example to illustrate the topic. Example 2.5 One period of the Fourier transform of x½n is depicted in Fig. 2.5. Draw the Fourier transform of y½n ¼ x½2n, i.e., draw Yn ðwÞ.

w

X n (w) 1

w

3

3

Fig. 2.5 One period of the Fourier transform of x½n

2.1 Sampling Rate Reduction by an Integer Factor …

79

Solution 2.5 First let’s draw the graph of Y1n ðwÞ ¼ 12 Xn ðw2 Þ. For this purpose, we multiply the frequency axis of Xn ðwÞ by 2 and vertical axis of Xn ðwÞ by 12. The resulting graph is shown in Fig. 2.6. In the second step, we shift the graph of Y1n ðwÞ to the left and right by multiples of 2p and P sum the shifted graphs. In other words, we draw the graph of Yn ðwÞ ¼ 1 k¼1 Y1n ðw  k2pÞ. The shifted graphs and their summation result are depicted in Figs. 2.7, 2.8, and 2.9. Right Shifted Functions: Left Shifted Functions: Sum of the Shifted Functions: Exercise: One period of the Fourier transform of x½n is depicted in Fig. 2.10. Draw the Fourier transform of y½n ¼ x½3n, i.e., draw Yn ðwÞ.

Y1n ( w)

w 1 Xn( ) 2 2

2

w

2

1 2

2

2 3

2

2 3

Fig. 2.6 The graph of 12 Xn ðw2 Þ

Y1n ( w 2 )

1 2

2

2

2 3

1 w 2 ) Xn( 2 2

2

2

w

2 3

Fig. 2.7 Right shifted functions

Y1n ( w 2 )

2

2 3

1 w 2 ) Xn( 2 2

2

2

Fig. 2.8 Left shifted functions

2 3

1 2

2

w

80

2 Multirate Signal Processing

Yn (w) 1 2

2

2 3

2

2

2 3

2 3

2 3

2 3

2

2

2

w

2 3

Fig. 2.9 Sum of the shifted functions

Fig. 2.10 One period of the Fourier transform of x½n

w

X n (w) 1

w

3

2.1.3

3

Aliasing in Downsampling

A digital signal is nothing but a mathematical sequence obtained via sampling of a continuous time signal. If we have sufficient number of samples, we can reconstruct the continuous time signal from its samples. If we have too many samples, generated during the sampling operation we can eliminate some of these excessive samples via the downsampling operation. However, while performing the downsampling operation, we should be careful to keep sufficient number of samples in the digital signal such that the reconstruction of the continuous time signal is still possible after downsampling operation. If we eliminate a number of samples more than a threshold value, the rest of the samples may not be sufficient to reconstruct the continuous time signal and this effect is seen as the aliasing in the spectrum graph of the downsampled signal. Example 2.6 Assume that we have a low pass continuous time signal with bandwidth fN ¼ 40 Hz. We choose the sampling frequency according to the criteria fs [ 2fN ! fs [ 80 as fs ¼ 120. This means that we take 120 samples per-second from the continuous time signal. However, our chosen sampling frequency is not very cost efficient. The lower limit for the sampling frequency is fs [ 80 which means that the minimum sampling frequency can be chosen as fs ¼ 81: However we use fs ¼ 120 which means that every per-second we transmit 120 − 81 = 39 excessive samples which are not necessary to reconstruct the continuous time signal. We can

2.1 Sampling Rate Reduction by an Integer Factor …

81

reconstruct the continuous time signal using only 81 samples. We can omit the excessive 39 samples via downsampling operation. Let’s now determine the criteria for no aliasing in downsampling operation. After downsampling operation, we have Mfs remaining samples per-second. If this number of remaining samples is greater than 2fN , then no aliasing occurs. That is if fs fs [ 2fN ! M\ M 2fN

ð2:21Þ

is satisfied, then aliasing is not seen in the spectrum of the downsampled signal. Let’s simplify (2.21) more as M\

fs 1 ! M\ 2 Ts fN 2fN |{z}

ð2:22Þ

fD

where fD is the digital frequency, and manipulating more, we have M\

1 p p ! M\ ! M\ ! MwD \p 2fD 2pfD wD

ð2:23Þ

where wD is the angular digital frequency. Let’s now graphically illustrate the no aliasing criteria after downsampling operation. Assume that one period of the Fourier transform of the digital signal x½n to be downsampled is given as in Fig. 2.11. Let y½n ¼ x½Mn be the downsampled signal. w Depending on the value of M, we can draw the two possible graphs of M1 Xn M as shown in Figs. 2.12 and 2.13. When the graph in Fig. 2.12 is shifted to the left and right by multiples of 2p, no overlapping occurs among shifted graphs. However, this case does not hold for the graph shown in Fig. 2.13. If the graph shown in Fig. 2.13 is shifted to the left and right by multiples of 2p, overlapping is observed between shifted replicas, and this situation is depicted in Fig. 2.14. Example 2.7 The continuous time signal xc ðtÞ ¼ cos ð6000ptÞ is sampled with 1 and the digital sequence x½n is obtained. Next the digital sampling period Ts ¼ 8000

w

X n (w)

Fig. 2.11 One period of the Fourier transform of the digital signal x½n to be downsampled

1

wD

wD

w

82

2 Multirate Signal Processing

X n (w / M )

M

w M

1 M

MwD

M Fig. 2.12 Case-1: Graph of

w

MwD

M



1 w M Xn M

X n (w / M )

M

w M

1 M

MwD

MwD

M

Fig. 2.13 Case-2: Graph of

w

M



1 w M Xn M

Yn (w) 1 M

2

MwD

MwD

2

w

Fig. 2.14 Aliasing in downsampled signal spectrum graph

signal x½n is downsampled and y½n ¼ x½4n is obtained. Decide whether aliasing occurs in spectrum of y½n or not. Solution 2.7 If the given continuous time signal is compared to cos ð2pftÞ, the frequency of the continuous time signal is found as f ¼ 3000 Hz. And the sampling frequency is fs ¼ 8000 Hz. After downsampling operation sampling frequency reduces to fs ¼ 8000 4 ¼ 2000 Hz and this value is less than 2f ¼ 6000 Hz. This means that aliasing is seen in the spectrum of y½n. 1 , and Exercise: For the system in Fig. 2.15, xc ðtÞ ¼ cos ð5000ptÞ, Ts ¼ 10;000 M ¼ 2. According to given information, draw the Fourier transforms of the signals xc ðtÞ; x½n; y½n, and yr ðtÞ, and also write the time domain expression for yr ðtÞ.

2.1 Sampling Rate Reduction by an Integer Factor …

83

y[n]

C/D

xc (t )

x[ Mn]

x[n]

yr (t )

D/C

M

Ts

Ts

Fig. 2.15 Signal processing system for exercise

2.1.4

Interpretation of the Downsampling in Terms of the Sampling Period

If x½n ¼ xc ðnTs Þ, then for the downsampled signal y½n ¼ x½Mn ! y½n ¼ xc ðn MTs Þ new sampling period is Ts0 ¼ MTs which is an integer multiple of Ts . The |{z} Ts0

digital signal obtained from xc ðtÞ using sampling period Ts is shown in Fig. 2.16. The digital signal x½n in Fig. 2.16 is written as a mathematical sequence as x½n ¼ ½   a

b c

d

e

f

g |{z}

h i

j k

l

m   :

n¼0

Now consider y½n ¼ x½2n ! y½n ¼ xc ðn2Ts Þ, in this case the samples are taken from xc ðtÞ at every Ts0 ¼ 2Ts . This operation is illustrated in Fig. 2.17. The digital signal y½n in Fig. 2.17 can be written as a mathematical sequence as y½n ¼ ½   a

c e

g |{z}

i

k

m   :

n¼0

Similarly, if g½n ¼ x½4n ! g½n ¼ xc ðn4Ts Þ, the samples are taken from xc ðtÞ at every Ts0 ¼ 4Ts . This operation is illustrated in Fig. 2.18.

c b

d

x c (t )

e

j i

f

a

− 6Ts − 5Ts − 4Ts − 3Ts− 2Ts − Ts

g

h

k l m

0 Ts 2Ts 3Ts 4Ts 5Ts 6Ts

Fig. 2.16 Sampling of the continuous time signal with sampling period Ts

t

84

2 Multirate Signal Processing

c

e

xc (t )

k i

a

6Ts

g

2Ts

4Ts

m

2Ts

0

4Ts

6Ts

t

Fig. 2.17 Sampling of the continuous time signal with sampling period 2Ts

xc (t )

c

k

g

4Ts

4Ts

0

t

Fig. 2.18 Sampling of the continuous time signal with sampling period 4Ts

The digital signal g½n in Fig. 2.18 can be written as a mathematical sequence as y½n ¼ ½   c

g |{z}

k   :

n¼0

Example 2.8 For the signal processing system given in Fig. 2.19, xc ðtÞ ¼ 1 , and M ¼ 3. Using the given information, calculate and cosð5000ptÞ, Ts ¼ 8000 draw the Fourier transforms of the signals xc ðtÞ; x½n; y½n, and yr ðtÞ. Besides, write the time domain expression for yr ðtÞ.

y[n]

xc (t )

C/D

x[n] M

Ts Fig. 2.19 Signal processing system for Example 2.8

x[ Mn]

D/C

Ts

yr (t )

2.1 Sampling Rate Reduction by an Integer Factor …

85

Solution 2.8 Before starting to the solution, let’s provide some background information as CosðhÞ ¼



 1  jh e þ ejh FT ejw0 t ¼ 2pdðw  w0 Þ 2

ð2:24Þ

FT fcosðwN tÞg ¼ pðdðw  wN Þ þ dðw þ wN ÞÞ:

ð2:25Þ

Accordingly, the Fourier transform of xc ðtÞ is found as Xc ðwÞ ¼ pðdðw  5000pÞ þ dðw þ 5000pÞÞ: and graphically it is shown in Fig. 2.20. For the given example, since fs [ 2fN ! 8000 [ 2  2500 criteria is satisfied, no aliasing is observed in the Fourier transform of x½n, and for this reason, one period of the Fourier transform of x½n for the interval p  w\p equals Xn ðwÞ ¼   1 w Ts Xc Ts which is depicted in Fig. 2.21. For the downsampled signal, we have y½n ¼ x½3n, let’s draw one period of  Yn ðwÞ ¼ 13 Xn w3 using one period of Xn ðwÞ around origin as in Fig. 2.22 where impulses are labeled with letters so that we can distinguish them while forming the Fourier transform of y½n. If the graph in Fig. 2.22 is carefully inspected, we see that after downsampling operation one period of the Fourier transform of the downsampled signal extends beyond the interval ðp; pÞ in frequency axis. This means that the number of samples omitted is greater than the allowed threshold and for this reason perfect reconstruction of the continuous time signal is not possible anymore. It may be reconstructed with some distortion or the reconstructed signal may be a totally X c (w)

Fig. 2.20 Fourier transform of xc ðtÞ in Example 2.8

w

0

5000

5000

w

X n (w)

Fig. 2.21 One period of the Fourier transform of x½n for Example 2.8

8000

5 8

0

5 8

w

86

2 Multirate Signal Processing

1 w Xn( ) 3 3 8000 3

A

15 8

B

15 8

0

Fig. 2.22 The graph of 13 Xn

8000 3

w 3

w

for Example 2.8

Br

Ar

2

8

31 8

w

Fig. 2.23 One period of Yn ðwÞ shifted to the right by 2p

different one. The amount of distortion in the reconstructed continuous time signal depends on the rate of the omitted samples, i.e., rate of the compression or rate of the downsampling. As the number of omitted samples increases, the amount of distortion in the reconstructed signal increases, as well. To get the graph of Yn ðwÞ, we shift its one period depicted in Fig. 2.22 to the left and to the right by multiples of 2p and sum the shifted replicas. The right shifted graph by 2p is given in Fig. 2.23. And the left shifted graph by 2p is shown in Fig. 2.24a. Summing the centered, right shifted, and left shifted graphs, we get the graph of Yn ðwÞ as shown in Fig. 2.24b. Now let’s find the expression for the reconstructed signal yr ðtÞ. For this purpose, we consider the graph of Yn ðwÞ for the interval p  w\p and draw Yr ðwÞ ¼ Ts Xn ðTs wÞ. To achieve this, we divide the frequency axis by Ts and multiply the amplitudes by Ts . These operations generate the graph depicted in Fig. 2.25. If the inverse Fourier transform of Yr ðwÞ depicted in Fig. 2.25 is calculated, we obtain the time domain expression of the reconstructed signal as 1 yr ðtÞ ¼ cos ð1000ptÞ 3

2.1 Sampling Rate Reduction by an Integer Factor …

87

(a) Bl

Al

8000 3

w

2

31 8

8

(b)

Yn (w) 8000 3

AI

31 8

A

BI

Ar

15 8

8

8

B

Br

15 8

31 8

Fig. 2.24 a One period of Yn ðwÞ shifted to the left by 2p. b The graph of Yn ðwÞ for Example 2.8

Yr (w)

Fig. 2.25 Fourier transform of the reconstructed signal for Example 2.8

B

1000

3

A

1000

w

which is quite different from the sampled signal xc ðtÞ ¼ cos ð5000ptÞ. The reason for this is that during the downsampling operation too many samples, beyond the allowable threshold, are omitted and this resulted in aliasing in frequency domain and perfect reconstruction of the original signal is not possible anymore. Question: During the downsampling operation we have to omit more samples than the number of allowable one. However, we want to decrease the effect of aliasing at the spectrum of the digital signal. What can we do for this? Answer: If y½n ¼ x½Mn alising occurs in Yn ðwÞ, if the largest frequency of Xn ðwÞ in the interval p  w\p is greater than Mp . This situation is depicted in Fig. 2.26. For the conversion of y½n to continuous time signal yr ðtÞ, the portion of Yn ðwÞ for the interval p  w\p in Fig. 2.26 is used. This portion is depicted alone in Fig. 2.27. As it is seen from Fig. 2.27, the overlapping shaded parts cause distortion in the reconstructed signal. Then how can we decrease the distortion amount? If we can

88

2 Multirate Signal Processing

M

M

w

M

MwD

M

1

1

wD

1 w Xn( ) M M

Y1n ( w)

w

X n (w)

M

w

wD

MwD

M

Yn ( w)

w

Y1n ( w k 2 ) k

1 M

2

MwD

MwD

w

2

Fig. 2.26 Aliasing case in downsampled signal Fig. 2.27 Yn ðwÞ, p  w\p

Yn (w)

w

1 M

w

eliminate the shaded regions in the spectrum of the downsampled signal, the reconstructed signal will have less distortion. However due to the clipping of the parts extending beyond the interval ðp; pÞ, some distortion will always be available in the reconstructed signal. This distortion is due to the information loss owing to the clipping of the spectrum regions in Fig. 2.26 for the intervals p  w\Mwd and Mp  w\p. What we do here is that we want try to decrease the amount of distortion, not complete elimination of it. Then if we can get a spectrum graph for Yn ðwÞ; p  w\p as shown in Fig. 2.28 the reconstructed signal will have less distortion. Fig. 2.28 After elimination of the overlapping shaded parts in Fig. 2.27

Yn (w)

w

1 M

w

2.1 Sampling Rate Reduction by an Integer Factor …

w

X n (w)

89

1

1

H dn (w)

wD

w

H dn ( w) X n ( w)

M

M

w

w

wD

M

M

Fig. 2.29 Elimination of the high frequency parts by a decimator filter

We can omit the overlapping shaded parts if we can filter high frequency portions of Xn ðwÞ before downsampling operation, i.e., the portions of Xn ðwÞ for the intervals Mp  w\p and p  w\  Mp should be filtered out. This can be achieved using a low pass filter as shown in bold lines Fig. 2.29. The lowpass filter clips the wigs of the signal that extends beyond the interval ðp; pÞ. And this clipping prevents the overlapping problem in downsampled signal spectrum. The lowpass filter used in Fig. 2.29 is called decimator filter whose frequency domain expression for its one period around origin is written as  Hdn ðwÞ ¼

1 0

if if

jwj\ Mp p M \jwj\p:

ð2:26Þ

The time domain expression of the decimator filter can be computed using the inverse Fourier transform as 1 hdn ½n ¼ 2p

Z w;2p

p

1 Hdn ðwÞejwn dw ! hdn ½n ¼ 2p

ZM 1  ejwn dw

ð2:27Þ

Mp

yielding the expression  n sin pn 1 M ! hdn ½n ¼ sin c hdn ½n ¼ : pn M M

ð2:28Þ

The filtering process before downsampling operation is illustrated in Fig. 2.30. The system in Fig. 2.30 is called decimator system, and the overall operation in Fig. 2.30 is named as decimation. For the system in Fig. 2.30, we have Y1n ðwÞ ¼ Hdn ðwÞXn ðwÞ and y½n ¼ y1 ½Mn. w Þ; p  w\p. One period of One period of Yn ðwÞ is written as Yn ðwÞ ¼ M1 Y1n ðM Yn ðwÞ is shown in Fig. 2.31.

90

2 Multirate Signal Processing H dn (w)

Fig. 2.30 Decimator system

1

w M

M

hd [n]

x[n]

y1[n]

Fig. 2.31 One period of Yn ðwÞ

M

y[n]

Yn (w )

w

y1[ Mn]

1 M

w

One period of Yn ðwÞ can be expressed as  Ynop ðwÞ ¼

Yn ðwÞ 0

p  w\p otherwise

ð2:29Þ

which can be used for the calculation of the Fourier transform of y½n as Yn ðwÞ ¼

1 X

Ynop ðw  k2pÞ:

ð2:30Þ

k¼1

Considering Fig. 2.31 the graph of (2.30) can be drawn as in Fig. 2.32. Exercise: If y½n ¼ x½3n and the Fourier transform of x½n for p  w\p is as given in Fig. 2.33, draw the Fourier transform of y½n, i.e., draw Yn ðwÞ. Downsampling can also be used for de-multiplexing operations, i.e., separating digital data to its components. We below give some examples to illustrate the use of downsampling for de-multiplexing operations. Yn (w) 1 M

2

Fig. 2.32 Fourier transform of filtered and dowsampled signal

2

w

2.1 Sampling Rate Reduction by an Integer Factor …

91

w

X n (w)

Fig. 2.33 One period of the Fourier transform of a digital signal

1

2 3

w

2 3

Note: The simplest de-multiplexer is the serial to parallel converter. Example 2.9 The delay system is described in Fig. 2.34. If x½n ¼ ½1

2 3

4

5

6 7

8

9 |{z}

10

11

12 13

14

15

n¼0

find the output of each unit given in Fig. 2.35. Solution 2.9 To get y½n ¼ x½n  n0 ; n0 [ 0, it is sufficient to shift n ¼ 0 pointer to the left by n0 units in x½n sequence. For negative n0 , we shift the n ¼ 0 pointer to the right by n0 units. According to this information, x½n  1 can be calculated as x½n  1 ¼ ½1

2 3

4

5

6 7

8 |{z}

9

10

11

12 13

14

15:

n¼0

If we divide the time axis by 2 and take only the integer division results, we get the signals y1 ½n ¼ ½1 3

5 7

9 |{z}

11

13

15 y2 ½n ¼ ½2

4 6

8 |{z}

n¼0

10

12

14

n¼0

at the outputs of the downsamplers. As it is seen from the obtained sequences, the system separates the odd and even indexed samples. Fig. 2.34 Delay system

Fig. 2.35 Signal processing system for Example 2.9

x[n]

z

x[n]

z

1

x[n n0 ]

n0

2

y1[n]

2

y2 [ n ]

92

2 Multirate Signal Processing

Fig. 2.36 Delay system

x[n]

Fig. 2.37 Signal processing system for Example 2.10

z

n0

x[n]

x[n n0 ]

3

y1[n]

z

1

3

y2 [ n ]

z

1

3

y3[n]

Example 2.10 The delay system is shown in Fig. 2.36. If x½n ¼ ½1

2 3

4

5

6 7

8

9 |{z}

10

11

12 13

14

15

n¼0

find the output of each unit given in Fig. 2.37. Solution 2.10 Following similar steps as in the previous example, we find the digital signals at the outputs of the downsamplers as y1 ½n ¼ ½ 3 y3 ½n ¼ ½ 1

6

9 12

4

7 10

15  13 

y2 ½n ¼ ½ 2

5

8 11

14 

which are nothing but sub-sequences obtained by dividing data signal x½n into non-overlapping sequences.

2.1.5

Drawing the Fourier Transform of Downsampled Signal in Case of Aliasing (Practical Method)

Let y½n ¼ x½Mn be the downsampled digital signal. To draw the Fourier transform of y½n in case of aliasing, we follow the subsequent steps. w Step 1: First we draw the graph of M1 Xn M . For this purpose, we divide the 1 horizontal axis of the graph of Xn ðwÞ by M , i.e., we multiply the horizontal axis by M, and multiply the amplitude values by 1=M.  w extends beyond the interval Step 2: In case of aliasing, the graph of M1 Xn M ðp; pÞ. The portion of the graph extending to the left of p is denoted by ‘A’, and the potion extending to the right of p is denoted by ‘B’.

2.1 Sampling Rate Reduction by an Integer Factor …

93

Step 3: The portion of the graph denoted by ‘A’ in Step 2 is shifted to the right by 2p, and the portion denoted by ‘B’ is shifted to the left by 2p. The overlapping lines are summed and one period of Yn ðwÞ around origin is obtained. Let’s denote this one period by Yn1 ðwÞ. Step 4: In the last step, one period of Yn ðwÞ around origin denoted by Yn1 ðwÞ is shifted to the left and right by multiples of 2p and all the shifted replicas are summed to get Yn ðwÞ, this is mathematically stated as

Yn ðwÞ ¼

1 X

Yn1 ðw  k2pÞ:

k¼1

Now let’s explain these steps using graphics. Let the Fourier transform of x½n be as shown in Fig. 2.38. w around origin will be as shown in In case of aliasing, one period of M1 Xn M Fig. 2.39. w If Fig. 2.39 is inspected carefully, it is seen that the function M1 Xn M takes values outside the interval ðp; pÞ on horizontal axis. In Fig. 2.40, the shadowed w extending outside of triangles denoted by ‘A’ and ‘B’ show the portion of M1 Xn M ðp; pÞ. X n (w)

Fig. 2.38 Fourier transform of x½n

A

wd

0

wd

w

1 w Xn( ) M M

Fig. 2.39 One period of 1 w X n M M around origin in case of aliasing

A M

Mwd

0

Mwd

w

94

2 Multirate Signal Processing

If the shadowed triangles ‘A’ and ‘B’ in Fig. 2.40 are shifted to the right and left by 2p, we obtain the graphic in Fig. 2.41. If the overlapping lines in Fig. 2.41 are summed, we obtain the graphic shown in bold lines in Fig. 2.42. As it is clear from Fig. 2.41, overlapping regions distorts the original signal. The amount of distortion depends  w on the widths of the shadowed triangles. In other words, as the function 1 X M n M extends outside the interval ðp; pÞ more, the amount of distortion on the original signal due to overlapping increases. The graph obtained after summing the overlapping lines is depicted alone in Fig. 2.43. 1 w Xn( ) M M

Fig. 2.40 One period of 1 w X n M M around origin in case of aliasing

A M

A

B

Mwd

Mwd

0

Fig. 2.41 Shaded parts shifted

w

A M

A

B Mwd

Mwd

0

Fig. 2.42 Sum of the overlapping lines

w

A M

B

Awd

A

0

Awd

w

2.1 Sampling Rate Reduction by an Integer Factor … Fig. 2.43 The resulting graph after summing the overlapping lines

95

A M

Mwd

Mwd

0

w

X n ( w)

Fig. 2.44 One period of Xn ðwÞ

1

3 8

0

3 8

w

Exercise 2.11 The Fourier transform of x½n, i.e., Xn ðwÞ, is shown in Fig. 2.44. Draw the Fourier transform of the downsampled signal y½n ¼ x½Mn; M ¼ 4. Solution 2.11 Step 1: First we draw the graph of



1 w M Xn M

as in Fig. 2.45.

For the graph of Fig. 2.45, the parts that fall outside of the interval ðp; pÞ are denoted by the shaded triangles ‘A’ and ‘B’ in Fig. 2.46. If the shaded parts ‘A’ and ‘B’ in Fig. 2.46 are shifted to the right and to the left by 2p, we obtain the graph in Fig. 2.47. The equations of the overlapping line on the interval ðp; p=2Þ in Fig. 2.47 1 1 1 w þ 14 and  12p w  24 , and when these equations are summed, can be written as 12p 5 we obtain 24. In a similar manner, the sum of the equations of the overlapping line 5 on the interval ðp=2; pÞ can be found as 24 . Hence one period of Yn ðwÞ around origin can be drawn as shown in Fig. 2.48. In the last step, shifting one period of Yn ðwÞ to the left and right by multiples of 2p and summing the shifted replicas we obtain the graph of Yn ðwÞ. 1 w Xn( ) 4 4

Fig. 2.45 One period of

1 w M Xn M

1/4

3 2

0

3 2

w

96

2 Multirate Signal Processing

Fig. 2.46 One period of

1 w Xn( ) 4 4

1 w M Xn M

1/4

A

B

0

3 2

Fig. 2.47 Shaded parts shifted to the right and to the left by 2p

w

3 2

1/4

B

A

0

2

2

Fig. 2.48 One period of Yn ðwÞ around origin

w

1/4

w

0

2

2

X n ( w)

Fig. 2.49 One period of the Fourier transform of x½n

4 2 1

3 8

0

4

4

3 8

Exercise: One period of the Fourier transform of x½n is shown in Fig. 2.49. Draw the Fourier transform of the downsampled signal y½n ¼ x½4n. Exercise: One period of the Fourier transform of x½n is shown in Fig. 2.50. Draw the Fourier transform of the downsampled signal y½n ¼ x½8n.

w

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

97

X n (w)

Fig. 2.50 Xn ðwÞ bir periyodu

1

3 16

2.2

0

3 24

w

Upsampling: Increasing the Sampling Rate by an Integer Factor

Assume that we want to transmit an analog signal. For this purpose, we first take some samples from the continuous time signal and form a mathematical sequence, and this process is called sampling. To decrease the transmission overhead, we omit some of the digital samples and this process is called downsampling. After downsampling operation, we transmit the remaining samples. At the receiver side, for better reconstruction of the analog signal, we try to find a method to increase the number of digital samples. For this purpose, we try to find the samples omitted during the downsampling operation. After finding the omitted samples, we can reconstruct the analog signal in a better manner. This means that first we reconstruct the original digital signal from downsampled digital signal then by using the reconstructed digital signal, we reconstruct the continuous time signal. Reconstruction of the original digital signal from the downsampled signal includes a two-step process. The first step is called up sampling also named as signal-expansion. In this step, the compressed signal, i.e., downsampled signal, is expanded in time axis, and for the new time instants, 0 values are assigned for the new amplitudes. The second step is called interpolation which is the reconstruction part for the omitted digital samples. In this part, the 0 values assigned to new time amplitudes for the expanded signal are replaced by the estimated values. Now let’s explain the upsampling operation.

2.2.1

Upsampling (Expansion)

The block diagram of the upsampler (expander) is shown in Fig. 2.51. The mathematical expression of the upsampling operation is Fig. 2.51 Upsampling operation

x[n]

L

y[ n ]

x[ n / L ]

98

2 Multirate Signal Processing

 n x L n ¼ 0; L; 2L; . . . y ½ n ¼ 0 otherwise:

ð2:31Þ

For simplicity of the expression we will assume that for the new time indices in the expanded signal, the amplitude values are 0, so we will not always explicitly write the second condition in (2.31), i.e., we will only use y½n ¼ x Ln to describe the signal expansion. To draw the graph of y½n ¼ x Ln , or to obtain the expanded signal, y½n ¼ x Ln we divide the time axis of x½n by 1=L, i.e., we multiply the time axis of x½n by L. This operation is illustrated with an example now. 13 15 17 find y½n ¼ x n3 . Example 2.12 If x½n ¼ ½1 3 5 7 9 |{z} 11 n¼0

Solution 2.12 The indices for amplitude values of x½n are explicitly written in x½n ¼ ½|{z} 1

3 |{z}

5 |{z}

7 |{z}

9 |{z}

11 |{z}

13 |{z}

15 |{z}

n¼5

n¼4

n¼3

n¼2

n¼1

n¼0

n¼1

n¼2

17 : |{z} n¼3

Dividing the indices of x½n by 1=3, i.e., multiplying the indices by 3, we get the sequence ½ |{z} 1

3 |{z}

5 |{z}

7 |{z}

9 |{z}

11 |{z}

13 |{z}

15 |{z}

n¼15

n¼12

n¼9

n¼6

n¼3

n¼0

n¼3

n¼6

17 : |{z} n¼9

Inserting missing indices and inserting 0 for amplitudes of the missing indices, we obtain the signal y½n as y½n ¼ ½1

0

0 3

0

0

5

0 0

7

0

0 9

0

0

11 |{z}

0 0

13

n¼0

0

2.2.2

0 15

0

0

17:

Mathematical Formulization of Upsampling

The upsampling, expansion, of x½n by L is defined as  n x L n ¼ 0; L; 2L; . . . y ½ n ¼ 0 otherwise

ð2:32Þ

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

99

which can be written in terms of impulse function as y ½ n ¼

1 X

x½kd½n  kL:

ð2:33Þ

k¼1

When the summation in (2.33) is expanded, we obtain y½n ¼    þ x½1d½n þ L þ x½0d½n þ x½1d½n  L þ    Note that to find x that is, if

n L , we simply insert L  1 zeros between two samples of x½n, x ½ n ¼ ½ a

b c

d

e ;

then to get x n4 simply insert 3 zeros between every two samples of x½n, and this operation yields x

2.2.3

h ni 4

¼ ½a

0

0 0

b

0

0

0 c

0

0

0 d

0

0

0

e:

Frequency Domain Analysis of Upsampling

Let’s try to find the Fourier transform of  n x L n ¼ 0; L; 2L; . . . y ½ n ¼ 0 otherwise:

ð2:34Þ

For this purpose, let’s start with the definition of the Fourier transform of y½n Yn ðwÞ ¼

1 X

y½nejwn

ð2:35Þ

n¼1

where substituting

P1 k¼1

x½k d½n  kL for y½n, we get

Yn ðwÞ ¼

1 1 X X

x½kd½n  kLejwn

ð2:36Þ

n¼1 k¼1

in which changing the order of summation terms, we obtain Yn ðwÞ ¼

1 1 X X k¼1 n¼1

x½kd½n  kLejwn

ð2:37Þ

100

2 Multirate Signal Processing

which can be rearranged as Yn ðwÞ ¼

1 X

x½k 

1 X

d½n  kLejwn

ð2:38Þ

n¼1

|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}

k¼1

ejwkL

yielding the expression 1 X

Yn ðwÞ ¼

x½k ejwkL :

ð2:39Þ

k¼1

If (2.39) is compared to the Fourier transform of x½n Xn ðwÞ ¼

1 X

x½nejwn

ð2:40Þ

n¼1

it is seen that Yn ðwÞ ¼ Xn ðLwÞ

ð2:41Þ

Referring to (2.41), it is understood that the graph of Yn ðwÞ can be obtained by dividing the frequency axis of Xn ðwÞ by L. As it is clear from (2.41) that the spectrum of the upsampled signal gets compressed in frequency domain. In fact, if a signal is expanded in time domain, it is compressed in frequency domain, similarly, if a signal is compressed in time domain, its spectrum expands in frequency domain. Example 2.13 One period of the Fourier transform of x½n around origin is given in

Fig. 2.52. Draw one period of the Fourier transform of y½n ¼ x Ln . Solution 2.13 Dividing the frequency axis of Xn ðwÞ by L, we obtain the Fourier transform of y½n which is depicted in Fig. 2.53. Note: Don’t forget that the Fourier transforms Xn ðwÞ and Yn ðwÞ are periodic functions with common period 2p. In fact, the Fourier transform of any digital signal is a periodic function with period 2p regardless whether the digital signal is periodic or not in time domain. If the digital signal is periodic in time domain then X n (w)

Fig. 2.52 One period of the Fourier transform of a digital signal

w

1

wD

wD

w

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

101

Yn (w)

Fig. 2.53 One period of the Fourier transform of upsampled signal for Example 2.12

1

L

w

wD L

wD L

L

w

L

L

its Fourier transform is an impulse train with period 2p, i.e., its Fourier transform is a discrete signal. Example 2.14 One period of the Fourier transform of x½n around origin is given in

Fig. 2.54. Draw one period of the Fourier transform of y½n ¼ x n2 . Solution 2.14 Dividing the frequency axis of Xn ðwÞ by 2, we get the graph in Fig. 2.55 for the Fourier transform of y½n. To get the graph in Fig. 2.55, we divided the horizontal axis of Xn ðwÞ by 2. Since Yn ðwÞ is a periodic function with period 2p, the graph in Fig. 2.55 can also be drawn for the interval p  w\p as shown in Fig. 2.56. Example 2.15 For the system given in Fig. 2.44 M ¼ L ¼ 2, and x½n ¼ ½|{z} 1

2

3 4

5

6

7 8

9

10:

n¼0

Find the signals xd ½n and y½n in Fig. 2.57. X n ( w)

Fig. 2.54 One period of the Fourier transform of a digital signal

w

1

w

2 3

2 3

Yn (w)

Fig. 2.55 One period of the Fourier transform of upsampled signal for Example 2.13

1

2

w

2

w 2

3

3

2

102

2 Multirate Signal Processing

Yn (w)

Fig. 2.56 One period of the Fourier transform of upsampled signal for Example 2.13

w

1

w 3

Fig. 2.57 Signal processing system for Example 2.14

xc (t )

3

x[n]

C/D

M

xd [n ]

y[n ] L

Ts

Solution 2.15 To find xd ½n, we divide the time indices of x½n by 2 and keep only integer division results. This operation yields 1 xd ½n ¼ ½|{z}

3

5

7 9:

n¼0

To find y½n, we divide the time indices of xd ½n by 12, i.e., multiply the time indices of xd ½n by 2. For new indices, amplitude values are equated to 0. The result of this operation is the signal y½n ¼ ½|ffl{zffl} 1 0

3

0

5

0 7

0

9:

n¼0

The overall procedure is illustrated in Fig. 2.58.

x[n] [ 1 2 3 4 5 6 7 8 9 10]

2

xd [ n] [ 1 3 5 7 9] n 0

n 0

2

y[n] [ 1 0 3 0 5 0 7 0 9 0] n 0

Fig. 2.58 Downsampling and upsampling

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

xc (t )

C/D

x[n]

M

xd [n ]

y[n ] L

103

D/C

y r (t )

Ts Fig. 2.59 Signal processing system

2.2.4

Interpolation

Let’s consider the signal processing system shown in Fig. 2.59. The system includes one downsampler, one upsampler and one D/C converter. Let’s study the reconstructed signal yr ðtÞ. Assume that y½n is a causal signal. The signal yr ðtÞ is calculated from the digital signal y½n using yr ð t Þ ¼

1 X

y½nhr ðt  nTs Þ

ð2:42Þ

n¼1

where hr ðtÞ can either be ideal reconstuction filter, i.e., hr ðtÞ ¼ sincðt=Ts Þ or triangular approximated reconstruction filter, or any other approximated filter. When we expand the summation in (2.42), we see that some of the shifted filters are multiplied by 0, since some of the samples of y½n are 0. The expansion of (2.42) happens to be as yr ðtÞ ¼ y½0hr ðtÞ þ y½1hr ðt  Ts Þ þ y½2hr ðt  2Ts Þ þ y½3hr ðt  3Ts Þ þ    ð2:43Þ yielding yr ðtÞ ¼ 1  hr ðtÞ þ 0  hr ðt  Ts Þ þ 3  hr ðt  2Ts Þ þ 0  hr ðt  3Ts Þ þ    ð2:44Þ Multiplication of some of the shifted filters by 0 results in information loss in the reconstructed signal. Question: So how can we increase the quality of the reconstructed signal? Answer: If we can replace 0 values in the expanded signal y½n by their estimated values, yr ðtÞ expression in (2.44) will not include 0 multiplication terms and reconstructed signal becomes better. That is, x [ n] = [ { 1 2 3 4 5 6 7 8 9 10 ] n=0

y [n ] = [ { 1 0 3 0 5 0 7 0 9 0] n=0

Replace 0's by the estimated values of the omitted samples Omitted samples are 2, 4, 6, 8, 10

104

2 Multirate Signal Processing

So how can we find a method to find approximate values for the omitted samples of original signal x½n? If we can approximate omitted samples, we can replace 0’s in the expanded signal by the approximated values, then reconstruct the continuous time signal. The quality of the reconstructed signal will be better. We know that the amplitude values of a continuous time signal at time instants ti and ti þ 1 does not change sharply. Otherwise, it violates the definition of continuous time signal. For instance, the amplitude values of a continuous time signal for three time instants are given in Fig. 2.60. Hence for the omitted samples, we can make a linear estimation. Assume that L ¼ M ¼ 2, in this case, during the downsampling operation; we omit one sample from every other 2 samples. After upsampling operation, we have 0 in the place of omitted sample. We can estimate the omitted sample using the neighbor samples of the omitted sample. In Fig. 2.60, assume that after sampling operation, we obtain the digital signal [a b c], and in this case, downsampled signal can be calculated as ½a c. The expanded signal or upsampled signal becomes as ½ a 0 c  where 0 can be replaced by the estimated value a þ2 c. In general if there are L  1 zeros between two samples of the expanded signal, we can estimate the omitted samples drawing a line between the amplitudes of these two samples as illustrated in Fig. 2.61. The missing samples in Fig. 2.61. can be calculated using y ½ ni  ¼ b þ

ab ðnk þ L1  ni Þ; L

i ¼ k : k þ L  1:

ð2:45Þ

a

t0

b

c

t1

t2

t

Fig. 2.60 Amplitude values of a continuous time signal for three distinct time instants Estimated Values for Omitted Samples

a b 0 nk

0 nk

0 1

Fig. 2.61 Linear estimation of the missing samples

0 nk

n L 1

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

105

Let D ¼ ab L , when (2.45) is expanded for i ¼ k : k þ L  1; we get the amplitude vector ½b þ ðL  1ÞD

b þ ðL  2ÞD    b þ 2D

Example 2.16 Let x½n ¼ ½|{z} 1 x½3n y½n ¼ xd in y½n.

n 3

2

5

7

b þ D:

ð2:46Þ

9 10 10 find the signals xd ½n ¼

n¼0

and using linear estimation method, estimate the missing samples

Solution 2.16 To calculate the downsampled signal, we divide the time axis of x½n by 3 and keep only integer division results, and in a similar manner, to calculate the upsampled signal, we multiply the time axis of the downsampled signal by 3, and for the new time instants 0’s are assigned for amplitude values. The downsampled and upsampled signals can be calculated as xd ½n ¼ ½|{z} 1

7

10 y½n ¼ ½|{z} 1

n¼0

0

0

7

0 0

10:

n¼0

and these signals are graphically shown in Fig. 2.62. The missing samples in upsampled signal can be calculated using D¼

ab ; L

and

½b þ ðL  1ÞD

b þ ðL  2ÞD



b þ 2D

b þ D

For the first 2 missing samples D¼

17 ! D ¼ 2 3

n xd [ ] 3

xd [3n]

x[n] 10 10

10

10

9 7

7

7

5

2 1

0 1 2 3 4 5 6

1

1

n

0 1 2 3 4 5 6

n

Fig. 2.62 Original signal, downsampled signal, upsampled signal

0 1 2 3 4 5 6

n

106

2 Multirate Signal Processing

and the missing samples are ½7 þ 2ð2Þ

7 þ 1ð2Þ ! ½3

4:

For the next 2 missing samples D¼

7  10 ! D ¼ 1 3

and the missing samples are ½10 + 2ð1Þ

10 þ 1ð1Þ ! ½8

9:

The calculation of the missing samples is graphically illustrated in Fig. 2.63. Hence with the estimated values, the upsampled signal becomes as y½n ¼ ½|{z} 1

3

4

7

8

ð2:47Þ

9 10:

n¼0

The original sequence before downsampling operation was x½n ¼ ½|{z} 1

2

5

7

9

10

ð2:48Þ

10:

n¼0

When (2.47) is compared to (2.48), we see that the calculated samples are close to the original omitted samples. Fig. 2.63 Calculation of the missing samples

Estimated omitted samples

9

10

8

7

4

3

1 0

1 2 3 4 5 6

n

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

2.2.5

107

Mathematical Analysis of Interpolation

We explained an estimation method for the calculation of missing samples in expanded signal. However, we did not follow a mathematical analysis. How can we find the missing samples in upsampled (expanded) signal using a mathematical approach? In time domain, it is difficult to find a mathematical approach for the estimation of missing samples. Let’s approach to the problem in frequency domain. Let’s consider the system involving downsampling and upsampling operations given in Fig. 2.64 where we assume that L ¼ M. Let’s assume that the Fourier transform of x½n is as in Fig. 2.65. We will inspect the Fourier transforms of y½n and x½n in Fig. 2.64 and find a relation between them. Considering Fig. 2.65 the Fourier transform of xd ½Mn can be drawn as in Fig. 2.66.

xc (t )

C/D

x[n]

xd [n]

M

L

y[n]

yr (t )

D/C

Ts Fig. 2.64 Signal processing system including upsampling and downsampling operations

X n (w) 1

2

0 M

M

2

w

Fig. 2.65 Fourier transform of a digital signal

X nd (w) 1 M

2 Fig. 2.66 Fourier transform of xd ½Mn

0

w 2

108

2 Multirate Signal Processing

Dividing the horizontal axis of the graph in Fig. 2.66 by L, we obtain the graph of Yn ðwÞ as Fig. 2.67. If we compare the graph of Xn ðwÞ in Fig. 2.65 to the graph of Yn ðwÞ in Fig. 2.67, it is seen that for pL  jwj\2p  pL Xn ðwÞ ¼ 0 but Yn ðwÞ 6¼ 0, and for other frequency intervals, Yn ðwÞ ¼ M1 Xn ðwÞ. This is illustrated in Fig. 2.68. How can we make Yn ðwÞ to be equal to Xn ðwÞ for all frequency values? This is possible if we multiply Yn ðwÞ by a lowpass digital filter with the transfer function as in Fig. 2.69. Since L ¼ M and Yi ðwÞ ¼ Hi ðwÞYn ðwÞ, we can show the multiplication of Hi ðwÞYn ðwÞ as in Fig. 2.70. The result of the above multiplication is depicted in Fig. 2.71. For L ¼ M; we have Yi ðwÞ ¼ Xn ðwÞ which means that yi ½n ¼ x½n, that is omitted samples are reconstructed perfectly. Let’s now do the time domain analysis of this reconstruction process. If Yi ðwÞ ¼ Hi ðwÞYn ðwÞ, then yi ½n ¼ hi ½n  y½n. The time domain expression hi ½n can be obtained via inverse Fourier transform 1 hi ½ n ¼ 2p

Z Hi ðwÞejwn dw

ð2:49Þ

2p

Yn (w) 1 M

2

2 L

0 L

L

2 L

w

2

Fig. 2.67 Fourier transform of the signal y½n in Fig. 2.64

1 M

2

2 L

0 L

L

2 L

These regions are not available in X n (w)

Fig. 2.68 Comparison of Xn ðwÞ and Yn ðwÞ

w

2

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

109

H i (w) L

w

0

2 L

2

L

Fig. 2.69 Lowpass digital filter

Yi ( w)

H i ( w)Yn ( w) L 1 M

2

2 L

0 L

L

w

2 L

2

Fig. 2.70 The multiplication of Hi ðwÞYn ðwÞ

Yi (w)

1

w

0

2 L

2

L

Fig. 2.71 The graph of Yi ðwÞ ¼ Hi ðwÞYn ðwÞ

where using the frontiers  pL ; pL, we get p

1 hi ½ n ¼ 2p

ZL Le pL

jwn

dw ! hi ½n ¼

sin

pn

pn L

L

ð2:50Þ

110

2 Multirate Signal Processing hi [ n ]

1

4L

3L

2L

L

L

2L

0

3L

4L

n

Fig. 2.72 The graph of sin cðn=LÞ

which can be expressed in terms of sin cðÞ function as hi ½n ¼ sin c

 n L

:

ð2:51Þ

The graph of sin cðn=LÞ is depicted in Fig. 2.72.  As it is seen from Fig. 2.72 that hi ½n ¼ sin c Ln equals to 0 when n is a multiple  of L. The digital filter with impulse response hi ½n ¼ sin c Ln is called interpolating filter which is used to reconstruct those digital samples omitted during downsampling operation, i.e., used to reconstruct missing samples in the expanded, or upsampled signal. Exercise: The continuous time signal xc ðtÞ ¼ cosð2ptÞ is sampled with sampling period Ts ¼ 1=8 s: (a) For a mathematical sequence x½n from the samples taken from continuous time signal in the interval 0–1 s. (b) x½n is downsampled by M ¼ 2, and xd ½n is the downsampled signal, find xd ½n. (c) The downsampled signal xd ½n is upsampled and let y½n be the upsampled signal, find y½n. (d) Calculate the missing samples in y½n using the ideal interpolation filter.

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

2.2.6

111

Approximation of the Ideal Interpolation Filter

Since digital sin cðÞ filter is an ideal filter, it is difficult to implement such filters, instead we can use an approximation of this digital filter. As it is clear from Fig. 2.72, the digital sin cðÞ filter includes a large main lobe centered upon origin, and many other side lobes. To approximate the digital sin cðÞ filter, we can use triangles for the lobes in Fig. 2.72. The simplest approximation is to use an isosceles triangle for the main lobe and omit the other side lobes. The simplest approximated digital can filter can be obtained as shown in Fig. 2.73. Referring to Fig. 2.73 the approximated interpolation filter can mathematically be expressed as 8n if  L  n\0 < L þ 1; hai ½n ¼  Ln þ 1; if 0  n\L ð2:52Þ : 0; otherwise which can be expressed in more compact form as  hai ½n ¼

 jLnj þ 1; if  L  n\L 0; otherwise:

ð2:53Þ

hi [n]

1

hai [n ]

4L

3L

2L

L

2L

0 L

Fig. 2.73 Approximation of the ideal interpolation filter

3L

4L

n

112

xc (t )

2 Multirate Signal Processing

C/D

Ts

hd [n] Decimator Filter

x[n]

M

xd [n]

Downsampler Compressor

L

y[n]

Upsampler Expander

hi [n]

yi [n]

D/C

yr (t )

Interpolation Filter

Used to reconstruct the samples omitted during downsampling operation

Used to prevent the aliasing after downsampling operation

Fig. 2.74 Signal processing system with interpolation filter

With the interpolation filter our complete signal processing system becomes as in Fig. 2.74. For the reconstruction of the samples omitted during downsampling operation, if approximated interpolating filter is used, the reconstructed digital signal can be written as 1 X yi ½n ¼ hai ½n * y½n ! yi ½n ¼ y½khai ½n  k ð2:54Þ k¼1

where hai ½n denotes the approximated reconstruction filter, or interpolation filter. Now let’s try to write a relation between xd ½n and yi ½n given in Fig. 2.74. We know that y ½ n ¼

1 X

xd ½k d½n  kL:

ð2:55Þ

k¼1

When (2.53) is replaced into yi ½n ¼ hi ½n * y½n

ð2:56Þ

we get 1 X

yi ½n ¼ hi ½n *

xd ½k d½n  kL

ð2:57Þ

k¼1

which is simplified as y i ½ n ¼

1 X k¼1

xd ½k hi ½n  kL:

ð2:58Þ

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

113

When (2.58) is expanded, we get the explicit form of yi ½n as yi ½n ¼    þ xd ½1hi ½n þ L þ xd ½0hi ½n þ xd ½1hi ½n  L þ   

ð2:59Þ

Using the ideal interpolation filter, i.e., ideal reconstruction filter, hi ½ n ¼

sin

pn L

pn L

in (2.58), we can write the reconstructed digital signal as 1 X

y i ½ n ¼

 sin xd ½k 

k¼1



pðnkLÞ L

ð2:60Þ

pðnkLÞ L

or in terms of sin cðÞ function, we can write (2.60) as 

 n  kL y i ½ n ¼ xd ½k  sin c : L k¼1 1 X

ð2:61Þ

P Note: Digital reconstructed signal expression yi ½n ¼ 1 i ½n  kL is k¼1 xd ½k hP 1 quite similar to the analog reconstructed signal expression xr ðtÞ ¼ k¼1 x½k  hr ðt  kTs Þ. Example 2.17 For the system given in Fig. 2.75 L ¼ M ¼ 3 and x½n ¼ ½1 2 3 4. Find xd ½n; y½n; and yi ½n. Use approximated linear digital filter for hi ½n. Solution 2.17 For L ¼ M ¼ 3, if x½n ¼ ½1 y½n ¼ ½1 0 0 4. To find yi ½n we can use either y i ½ n ¼

1 X

2 3

4, then xd ½n ¼ ½1

4 and

y½khai ½n  k

ð2:62Þ

xd ½k hi ½n  kL

ð2:63Þ

k¼1

or y i ½ n ¼

1 X k¼1

Let’s use both of them separately. First using (2.53), let’s calculate and draw the linear approximated digital interpolation filter as in Fig. 2.76. Fig. 2.75 Signal processing system for Example 2.16

xd [n]

x[n] M

y[n] L

hi [n]

yi [n]

114

2 Multirate Signal Processing

hai [n]

Fig. 2.76 Approximated interpolation filter

|n|

hai [ n ] 3

1

n

3

1

1

3

3

2

3

2

2 3

3

1

3

1

1

0

2

3

n

Expanding (2.62), we get yi ½n ¼ y½0hai ½n þ y½1hai ½n  1 þ y½2hai ½n  2 þ y½3hai ½n  3:

ð2:64Þ

If y½n ¼ ½1 0 0 4 is considered, we see that the amplitude values at indices n ¼ 1; and n ¼ 2, are missing. When n ¼ 1 is placed into (2.64), we get yi ½1 ¼ y½0 hai ½1 þ y½1 hai ½0 þ y½2 hai ½1 þ y½3 hai ½2 |{z} |ffl{zffl} |{z} |ffl{zffl} |{z} |fflfflffl{zfflfflffl} |{z} |fflfflffl{zfflfflffl} 1

2=3

0

1

0

2=3

4

ð2:65Þ

1=3

which yields y i ½ 1 ¼

2 4 þ ! y i ½ 1 ¼ 2 3 3

ð2:66Þ

and when n ¼ 2 is placed into (2.64), we obtain yi ½2 ¼ y½0 hai ½2 þ y½1 hai ½1 þ y½2 hai ½0 þ y½3 hai ½1 |{z} |ffl{zffl} |{z} |ffl{zffl} |{z} |ffl{zffl} |{z} |fflfflffl{zfflfflffl} 1

1=3

0

1

0

2=3

4

ð2:67Þ

2=3

which yields y i ½ 2 ¼

1 8 þ ! y i ½ 2 ¼ 3 3 3

ð2:68Þ

So missing samples are found as yi ½1 ¼ 2 and yi ½2 ¼ 3, and when these samples are replaced by 0’s in y½n, we get yi ½n ¼ ½1

2

3

4

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

115

Now let’s use the formula 1 X

y i ½ n ¼

xd ½k hi ½n  kL:

ð2:69Þ

k¼1

When (2.69) is expanded, noting that xd ½n ¼ ½1

4 and L ¼ 3, we get

yi ½n ¼ xd ½0hai ½n þ xd ½1hai ½n  3:

ð2:70Þ

When (2.70) is evaluated for n ¼ 1, we obtain yi ½1 ¼ xd ½0 hai ½1 þ xd ½1 hai ½2 |ffl{zffl} |ffl{zffl} |ffl{zffl} |fflfflffl{zfflfflffl} 1

2=3

4

1=3

which yields y i ½ 1 ¼

2 4 þ ! y i ½ 1 ¼ 2 3 3

ð2:71Þ

and when (2.69) is evaluated for n ¼ 2, we get yi ½2 ¼ xd ½0 hai ½2 þ xd ½1 hai ½1 |ffl{zffl} |ffl{zffl} |ffl{zffl} |fflfflffl{zfflfflffl} 1

1=3

4

ð2:72Þ

2=3

which yields y i ½ 2 ¼

1 8 þ ! yi ½2 ¼ 3: 3 3

ð2:73Þ

Hence, both formulas give the same results. In addition, we had already introduced the linear estimation method using the continuity property of analog signals. It is now very clear that the linear estimation method is nothing but the use of triangle approximated digital reconstruction filter. Example 2.18 Show that the systems given in Fig. 2.77 have the same outputs for the same inputs.

Fig. 2.77 Signal processing systems for Example 2.17

x[n]

M

x[n]

H n (Mw)

xa [n]

xb [n]

H n (w)

y[n]

M

y[n]

116

2 Multirate Signal Processing

Solution 2.18 For the first system we have Xan ðwÞ ¼

  1 X 1M w  k2p Xn M k¼0 M

ð2:74Þ

and Yn ðwÞ ¼ Hn ðwÞXan ðwÞ ! Yn ðwÞ ¼

  1 X Hn ðwÞ M w  k2p Xn M k¼0 M

ð2:75Þ

For the second system we have Xbn ðwÞ ¼ Hn ðMwÞXn ðwÞ

ð2:76Þ

  1 X 1M w  k2p Xbn Yn ðwÞ ¼ : M k¼0 M

ð2:77Þ

and

When (2.76) is inserted into (2.77), we obtain X  w  k2p w  k2p 1 M1 Xn : Hn M Yn ðwÞ ¼ M k¼0 M M

ð2:78Þ

Since Hn ðwÞ is a periodic function with period 2p, (2.78) can be written as   1 X 1M w  k2p Yn ðwÞ ¼ Hn ðwÞXn M k¼0 M

ð2:79Þ

which is equal to X w  k2p 1 M1 ! Yn ðwÞ ¼ Hn ðwÞXan ðwÞ: Xn Yn ðwÞ ¼ Hn ðwÞ M k¼0 M

ð2:80Þ

When (2.75) is compared to (2.80), we see that both systems have the same outputs for the same inputs. Exercise: Show that the systems given below have the same outputs for the same inputs (Fig. 2.78). Example 2.19 For the system given in Fig. 2.79, find a relation in time domain between system input x½n and system output y½n.

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor Fig. 2.78 Signal processing system for exercise

Fig. 2.79 Signal processing system for Example 2.18

117

xa [n]

x[n]

L

x[n]

H n (w)

x[n]

L

xb [n]

xd [n]

H n (Lw)

L

L

y[n]

y[n]

y[n]

Solution 2.19 We have xd ½n ¼ x½Ln and y½n ¼ xd Ln . Putting xd ½n expression into y½n expression, we get y½n ¼ x½Ln L  ! y½n ¼ x½n. However, this is not always correct. Since we know that for L ¼ 2 if x½n ¼ ½1 2 3, then xd ½n ¼ ½1 3 and y½n ¼ ½1 0 3, it is obvious that x½n 6¼ y½n.

But using xd ½n ¼ x½Ln and y½n ¼ xd Ln ; we found y½n ¼ x½n. So, what is wrong with our approach to the problem? Because, we

did not pay attention to the criteria in upsampling operation. That is, y½n ¼ xd Ln if n ¼ kL; k 2 Z; otherwise, y½n ¼ 0. Then y½n ¼ x½n is valid only for some values of n and these n values are multiples of L. That is for L ¼ 2 if x½n ¼ ½1 2 3, then xd ½n ¼ ½1 3 and y½n ¼ ½1 0 3, and y½n ¼ x½n for n ¼ 0; 2 only. However, for some signals, no information loss occurs after compression operation. This is possible if the omitted samples are also zeros. In this case, expanded signal equals to the original signal. For example, if x½n ¼ ½|{z} a 0

b

0

c 0

b c

d

b

c 0

d

n¼0

then after downsampling by L ¼ 2, we get xd ½n ¼ ½a and after expansion by L ¼ 2, we obtain y½n ¼ ½|{z} a

0

0

d

n¼0

Thus, we see that y½n ¼ x½n for every n values. To write a mathematical expression between x½n and y½n, let’s express xd ½n in terms of x½n as

118

2 Multirate Signal Processing

xd ½n ¼

1 X

x½n

n¼1

1 X

d½n  rM

ð2:81Þ

r¼1

and express y½n in terms of xd ½n as 1 X

y ½ n ¼

xd ½k d½n  kL:

ð2:82Þ

k¼1

Inserting (2.81) into (2.82), we obtain y ½ n ¼

1 X k¼1

1 X

x½ k 

d½k  rM

r¼1

1 X

d½n  kL

ð2:83Þ

n¼1

which is the final expression showing the relation between x½n and y½n. Example 2.20 Find a method to check whether information loss occurs or not after downsampling by M. Solution 2.20 If x½n is downsampled by M, we omit M  1 samples from every M samples. If we denote the information bit indices by the numbers 0; 1; 2; . . .; M. . .; then the first omitted samples have indices 1; 2; . . .; M  1 and the second set of omitted indices have indices M þ 1; M þ 2; . . .; 2M  1, and so on. Hence, by summing the absolute values of the omitted samples and checking whether it equals to zero or not, we can conclude whether information loss occurs or not after downsampling operation. That is, we calculate Loss ¼

1 M 1 X X

jx½n þ kMj

ð2:84Þ

k¼1 n¼1

and if Loss 6¼ 0, then information loss occurs after downsampling of x½n, otherwise not. Example 2.21 If  y ½ n ¼

x ½ n 0

if n is even otherwise

ð2:85Þ

then write a mathematical expression between x½n and y½n. Solution 2.21 Using (2.85), we can express y½n in terms of x½n as y ½ n ¼

1 þ ð1Þn x½n: 2

ð2:86Þ

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor Fig. 2.80 Signal processing system for Example 2.21

xc (t )

C/D

119

xd [n]

x[n]

3

2

y[n]

Ts

Since cosðpnÞ ¼ ð1Þn , then (2.86) can also be written as y½n ¼

1 þ cosðpnÞ x½n: 2

1 Example 2.22 For the system given in Fig. 2.80, xc ðtÞ ¼ cosð2000ptÞ, Ts ¼ 4000 sec find x½n; xd ½n and y½n.

Solution 2.22 When continuous time signal is sampled, we get   p  1 ! x½n ¼ cos n : x½n ¼ xc ðtÞjt¼nTs ! x½n ¼ cos 2000pn 4000 2

ð2:87Þ

After downsampling operation, we have   3p xd ½n ¼ x½3n ! xd ½n ¼ cos n 2

ð2:88Þ

After upsampling operation, we have  y ½ n ¼

xd 0

n 2

n is even otherwise

ð2:89Þ

which yields  y½n ¼

cos 0

p n is even 4n otherwise

ð2:90Þ

The mathematical expression in (2.90) can be written in a more compact manner as y½n ¼

1 þ cosðpnÞ p  cos n : 2 4

ð2:91Þ

Using the property 1 cosðaÞ cosðbÞ ¼ ðcosða þ bÞ þ cosða  bÞÞ 2

ð2:92Þ

120

2 Multirate Signal Processing

Equation (2.91) can be written as     1 p  1 5p 1 3p n þ cos n y½n ¼ cos n þ cos 2 4 4 4 4 4

ð2:93Þ

where using cosðhÞ ¼ cosð2p  hÞ Eq. (2.93) can be written as   1 p  1 3p n : y½n ¼ cos n þ cos 2 4 2 4 Note: cos

ð2:94Þ

5p  5p 3p 5p 4 n ¼ cos 2pn  4 n ! cos 4 n ¼ cos 4 n

Example 2.23 xc ðtÞ ¼ ejwN t and x½n ¼ xc ðtÞjt¼nTs , Ts ¼ 1 find the Fourier transforms of xc ðtÞ and x½n. Solution 2.23 The Fourier transform of the continuous time exponential signal is Xc ðwÞ ¼ 2pdðw  wN Þ

ð2:95Þ

which is depicted in Fig. 2.81. If x½n ¼ xc ðtÞjt¼nTs , then one period of the Fourier transform of x½n is   1 w ; ð2:96Þ Xn ðwÞ ¼ Xc jwj\p Ts Ts which is shown in Fig. 2.82. Figure 2.82 can mathematically be expressed as Xn ðwÞ ¼ 2pdðw  wD Þ, jwj\2p. Since Xn ðwÞ is the Fourier transform of a digital signal, it is a periodic function and its period equals to 2p and it can be written as Xn ðwÞ ¼ 2p

1 X

dðw  wD  k2pÞ:

ð2:97Þ

k¼1

Fig. 2.81 Fourier transform of continuous time exponential signal

Xc (w)

2π

wN

0

Fig. 2.82 One period of the Fourier transform of digital exponential signal

X n (w)

w

|w|

2

0

wD

wN

w

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

121

After sampling of the continuous time exponential signal, we obtain j wN Ts n

x ½ n ¼ e

|fflffl{zfflffl}

! x½n ¼ ejwD n :

wD

Hence we can write the following transform pair in general FT

ejw0 n $ 2p

1 X

dðw  w0  k2pÞ:

ð2:98Þ

k¼1 p

Example 2.24 Given x½n ¼ ej3n , find Fourier transform of x½n, i.e., Xn ðwÞ: Solution 2.24 Xn ðwÞ ¼ 2pdðw  p3Þ, jwj\p and Xn ðwÞ is periodic with period 2p, so in more compact form, we can write it as Xn ðwÞ ¼ 2p

1 X k¼1

dðw 

p  k2pÞ 3

ð2:99Þ

Example 2.25 x½n ¼ cosðw0 nÞ, y½n ¼ cosðp3 nÞ, w½n ¼ cosð2p 3 nÞ, find the Fourier transforms of x½n; y½n; and w½n.   Solution 2.25 We know that cosðhÞ ¼ 12 ejh þ ejh and sinðhÞ ¼ 2j1 ejh  ejh , and using the Fourier transform of digital exponential function, we obtain the results Xn ðwÞ ¼ pðdðw  w0 Þ þ dðw þ w0 ÞÞ; jwj\p    p p Yn ðwÞ ¼ p d w  þd wþ ; jwj\p 3     3  2p 2p Wn ðwÞ ¼ p d w  þd wþ ; jwj\p: 3 3 Xn ðwÞ, Yn ðwÞ, and Wn ðwÞ are periodic functions with period 2p. Example 2.26 The transfer function of a lowpass digital filter is depicted in Fig. 2.84. Accordingly, find the output of the block diagram shown in Fig. 2.83 for the input signal   p  2p x½n ¼ cos n þ cos n : 3 3 The Fourier transform of the filter impulse is given as in Fig. 2.84. Fig. 2.83 Lowpass filtering of digital signals

x[n]

H n (w)

x f [n]

122

2 Multirate Signal Processing

H n (w)

Fig. 2.84 Digital lowpass filter transfer function

1

w

0

2

2

2

2

Solution 2.26 If digital frequency w is between  p2 and p2, that is if jwj\ p2, the digital frequency is accepted as low frequency. On the other hand, if p2 \jwj\p, the digital frequency is accepted as high frequency. One period of Fourier transform of x½n can be calculated as         p p 2p 2p þd wþ þp d w  þd wþ ; jwj\p Xn ðwÞ ¼ p d w  3 3 3 3 ð2:100Þ which is graphically illustrated in Fig. 2.85. At the output of the block diagram, we have Xfn ðwÞ ¼ Hn ðwÞXn ðwÞ and this multiplication is graphically illustrated in Fig. 2.86. As it is obvious from Fig. 2.86, the signal Xfn ðwÞ ¼ Hn ðwÞXn ðwÞ equals to    p p  Xfn ðwÞ ¼ p d w  þd wþ : 3 3

ð2:101Þ

X n (w) | w |

Fig. 2.85 Fourier transform of the input signal in Example 2.25

2 3 Fig. 2.86 Multiplication of Xn ðwÞ and Hn ðwÞ

w

0

2

3

3

2

2 3

2

2 3

X n (w) | w |

H n (w)

1

2 3

w

0 2

3

3

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

123

That is, high frequency part of the signal is filtered by the low pass filter, and at the output of the filter, only low frequency components exist. In time domain, the filter output equals to xf ½n ¼ cos

p  n : 3

ð2:102Þ

Example 2.27 In the system of Fig. 2.87, xc ðtÞ ¼ cosð2000ptÞ þ cosð5000ptÞ, Ts ¼ 1 3000 and transfer function of the digital filter is depicted in Fig. 2.88. Find x½n; xf ½n; and xd ½n. Solution 2.27 x½n ¼ xc ðtÞjt¼nTs leads to x½n ¼ cos Since cos becomes as

    2p 5p n þ cos n : 3 3

ð2:103Þ

5p  5p p 5p 3 n ¼ cos 2pn  3 n ! cos 3 n ¼ cos 3 n ,

x½n ¼ cos

then

  p  2p n þ cos n : 3 3

(2.103)

ð2:104Þ

The digital filter eliminates high frequency component of x½n, hence at the output of the filter we have xf ½n ¼ cos

Fig. 2.87 Signal processing system for Example 2.26

xc (t )

p  n : 3

C/D

x[n]

ð2:105Þ

H n (w)

x f [n] 2

xd [n]

Ts

H n (w)

Fig. 2.88 Digital lowpass filter transfer function

1

2

w

0

2

2

2

124

2 Multirate Signal Processing

x[n]

n0

z

x[n n0 ]

Fig. 2.89 Delay system

xa [n]

x[n]

xc [n]

M

xe [n]

M

xr [n]

xg [n] z

z

xb [n]

xd [n]

M

1

x f [n]

M

Fig. 2.90 Signal processing system for Example 2.27

After downsampling operation, we get xd ½n ¼ xf ½2n ! xd ½n ¼ cos

  2p n : 3

ð2:106Þ

Example 2.28 The delay system is shown in Fig. 2.89. In the system shown in Fig. 2.90, M ¼ 2 and x½n ¼ ½ 1 Find xa ½n; xb ½n; xc ½n; xd ½n; xe ½n; xf ½n and xr ½n. Solution 2.28 If x½n ¼ ½ 1

2

3

2

3 4

1 6 , then xa ½n ¼ ½|{z}

4 5

6 .

5 2

3

4

n¼0

5 6 and since xb ½n ¼ x½n þ 1 moving n ¼ 0 pointer to the right by one unit, we get xb ½n ¼ ½1

2 |{z}

3

4 5

6

n¼0

After downsampling, we have 1 xc ½n ¼ ½|{z}

3

5

xd ½n ¼ ½|{z} 2

n¼0

4

6:

0

4

n¼0

After upsampling, we have 1 xe ½n ¼ ½|{z}

0

3

0 5

xf ½n ¼ ½|{z} 2

n¼0

n¼0

After delay operator z1 , we have 0 xg ½n ¼ ½|{z} n¼0

2 0

4

0

6:

0

6:

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

125

And at the system output, we have xr ½n ¼ xe ½n þ xg ½n where 1 xe ½n ¼ ½|{z}

0 3

5 xg ½n ¼ ½|{z} 0

0

n¼0

2

0 4

0

6:

n¼0

Hence, x r ½ n ¼ ½ 1 2

3

4

5 6:

The signal flow of the system in Fig. 2.90 is shown in Fig. 2.91. Exercise: For the system given in Fig. 2.92, M ¼ 3 and x ½ n ¼ ½ 1 2

3

4

5 6

7

8

9 10

11

12

13

14 15:

Find the output of every block and finally find xr ½n. [ 1 0 3 0 5 0]

[ 1 3 5]

[ 1 2 3 4 5 6]

n 0

2

n 0

2

[ 1 2 3 4 5 6]

n 0

n 0

[ 0 2 0 4 0 6] n 0

z

1

[ 2 4 6] [1 2 3 4 5 6]

n 0

2

2

[ 2 0 4 0 6 0] n 0

n 0

Fig. 2.91 Signal flow for the system in Fig. 2.90

xa [n]

x[n]

M

xd [n]

M

xg [n]

xr [n] x j [n]

z

xb [n]

M

xe [n]

M

M

x f [n]

Fig. 2.92 Signal processing system for exercise

M

1

z

1

xh [n]

z

xc [n]

z

xi [n]

126

2 Multirate Signal Processing

x[n]

H 0 ( w)

xa [n]

M

xd [n]

G0 ( w)

M

xg [n]

xr [n]

x j [n]

z

H1 ( w)

xb [n]

M

xe [n]

G1 ( w)

M

1

xh [n]

Fig. 2.93 Signal processing system for Example 2.28

Example 2.29 For the system shown in Fig. 2.93, x½n ¼ cos M ¼ 2. Find H0 ðwÞ, H1 ðwÞ; G0 ðwÞ; and G1 ðwÞ such that xr ½n ¼ x½n.

2p p 3 n þ cosð3 nÞ,

Solution 2.29 H0 ðwÞ can be chosen as a low pass digital filter. H1 ðwÞ can be chosen as a high pass digital filter. G0 ðwÞ and G1 ðwÞ are interpolating sin cðÞ filters.

2.2.7

Anti-aliasing Filter

Consider the continuous to digital conversion system shown in Fig. 2.94. We know that to obtain one period the Fourier transform of x½n, we multiply the frequency axis of the Fourier transform of xc ðtÞ by Ts and multiply the amplitude 

Xc Tws . If the   Fourier transform of xc ðtÞ has a bandwidth greater than p=Ts , then T1s Xc Tws extends

axis of the Fourier transform of xc ðtÞ by 1=Ts , i.e., we calculate

1 Ts

beyond (p; p) and aliasing observed in the Fourier transform of x½n. This situation is described in Fig. 2.95.   Since Xn ðwÞ is periodic with period 2p when

1 Ts

Xc

w Ts

extends beyond (p; p),

overlapping will be observed in Xn ðwÞ as shown in Fig. 2.96. The portion of Xn ðwÞ in Fig. 2.96 for jwj\p is shown in Fig. 2.96. To decrease the effect of aliasing (overlapping) in the digital signal, we can filter the spectral components for jwj [ p=Ts in Xc ðwÞ before sampling operation. In this way, we can eliminate the overlapping shaded parts in Fig. 2.97. We name this filter as anti-aliasing filter and it is mathematically defined as

Fig. 2.94 Continuous to digital conversion

x c (t )

C/D

Ts

x [n]

2.2 Upsampling: Increasing the Sampling Rate by an Integer Factor

1

x[n]

0

Ts

w 1 Xc( ) | w | Ts Ts

X n ( w)

X c (w)

wN

127

wN

Ts

1 Ts

xc (nTs )

w

Ts wN

Ts wN

0

w

Fig. 2.95 Aliasing case in the Fourier transform of x½n

X n (w) 1 Ts

2

TswN

w

Ts wN 2

0

Fig. 2.96 Aliasing in Xn ðwÞ

Fig. 2.97 Xn ðwÞ in Fig. 2.96 for jwj\p

X n (w)

| w|

1 Ts

0

 Haa ðwÞ ¼

1 0

if jwj\ Tps otherwise

w

ð2:107Þ

whose time domain expression can be computed using inverse Fourier transform 1 haa ðtÞ ¼ 2p

Z1 Haa ðwÞejwt dw 1

128

2 Multirate Signal Processing

X c (w)

1

1

H aa (w)

wN

H aa ( w) X c ( w)

Ts

Ts

w

w

wN

Ts

Ts

Fig. 2.98 Anti-aliasing filtering X n (w) 1 Ts

2

2

w

Fig. 2.99 The Fourier transform of a digital signal obtained by sampling of a continuous time signal filtered by an anti-aliasing filter

as   sin haa ðtÞ ¼

pt Ts

pt

:

ð2:108Þ

Anti-aliasing filtering is shown in Fig. 2.98. The digital signal obtained after sampling of the filtered analog signal shown in Fig. 2.98 has the Fourier transform depicted in Fig. 2.99.

2.3

Practical Implementations of C/D and D/C Converters

Up to now we have studied theoretical C/D and D/C converter systems. However, the practical implementation of these units in real life shows some differences. The practical implementation of the C/D converter is shown in the first part of Fig. 2.100, and in a similar manner, the practical implementation of the D/C converter is shown in the second part of Fig. 2.100. C/D and D/C conversion systems include analog-to-digital and digital-to-analog converter units and the contents of these units are shown in Fig. 2.100. Now we will inspect every component of the complete system shown in Fig. 2.100.

2.3 Practical Implementations of C/D and D/C … Anti-Aliasing Analog to Digital Filter Conversion

Haa (w)

xc (t )

xa (t )

xa (t )

Sample and Hold

A/D

xo (t )

Digital Code

Quantization and Coding

Digital Signal Processing

129

Digital Code

Digital Code

Digital to Analog Conversion ^

D /A

Convert Digital Codes to Real Numbers

Digital Code

x o (t )

Reconstruction Filter

Zero Order Hold

xr (t )

^

x o (t )

Ts

Ts

Fig. 2.100 Practical implementations of C/D and D/C converter systems

2.3.1

C/D Conversion

A practical C/D converter includes the units shown in Fig. 2.101. Where antialiasing filter is used to decrease of amount of distortion in digital signal in case of aliasing. Antialiasing filter is defined as  Haa ¼

1 0

jwj\ Tps otherwise

ð2:109Þ

Inside A=D converter, we have Sample-and-Hold and Quantizer-Coder units which are shown in Fig. 2.102. For the coding of quantization levels, two’s complement, one’s complement or unsigned binary representations can be used. Once the analog signal is represented by bit sequences, i.e., codes, these bit sequences are processed depending on the application. For instance, in digital communication, these bit sequences are encoded by channel codes and obtained bit sequences are converted to complex symbols, i.e., digitally modulated, and transmitted. In data storage, these bit sequences are again coded using forward error corrections codes, such as Reed Solomon codes as in compact disc storage, and stored. Alternatively, these bit sequences can be passed through data compression algorithms and then stored. Anti-Aliasing Filter

Fig. 2.101 Practical C/D converter.

xc (t )

H aa (w)

Analog to Digital Conversion

xa (t )

A/ D

Digital Code

130

2 Multirate Signal Processing

Fig. 2.102 Components of A/D converter

xa (t )

xa (t )

Sample and Hold

Digital Code

A/ D

xo (t )

Quantization and Coding

Digital Code

Ts

2.3.2

Sample and Hold

The aim of the sample and hold circuit is to produce a rectangular signal and the amplitudes of the rectangles are determined at the sampling time instants. The simplest sample and hold circuit as shown in Fig. 2.103 which is constructed using a capacitor. Since usually sampling frequency fs is a large number, such as 10 kHz etc., it is logical to use a digital switch for the place of a mechanical switch as shown in Fig. 2.104. In the literature, much better sample and hold circuits are available. To give an idea about design improvement, the circuit in Fig. 2.104 can be improved by appending a buffer to the output preventing back current flows etc., and this improved circuit is shown in Fig. 2.105. The sample and hold operation for the input sine signal is illustrated in Fig. 2.106. Fig. 2.103 A simple sample and hold circuit

xo (t )

xc (t )

f s Hz

Fig. 2.104 Mechanical switch is replaced by an electronic switch

Ts

xc (t )

xo (t )

2.3 Practical Implementations of C/D and D/C …

131

f s Hz Ts

xc (t )

xo (t )

Fig. 2.105 Sample and hold circuit with a buffer at its output

xc (t )

Ts

2Ts 3Ts 4Ts 5Ts

xo (t )

6Ts

7Ts 8Ts 9Ts 10Ts

11Ts 12Ts 13Ts

t

Fig. 2.106 Calculation of the output of the sample and hold circuit for sine input signal

For sine input signal after sample and hold operation, we obtain the signal xo ðtÞ which is depicted alone in Fig. 2.107. Question: Can we write a mathematical expression for the signal xo ðtÞ shown in Fig. 2.107. Yes, we can write. For this purpose, let’s first define ho ðtÞ function as shown in Fig. 2.108. If the graph of xo ðtÞ in Fig. 2.107 is inspected, it is seen that xo ðtÞ signal is nothing but sum of the shifted and scaled ho ðtÞ functions. Using ho ðtÞ functions, we can write xo ðtÞ as xo ðtÞ ¼

1 X

xc ðnTs Þho ðt  nTs Þ ! xo ðtÞ ¼

k¼1

1 X

x½nho ðt  nTs Þ:

ð2:110Þ

k¼1

xo (t )

Ts

2Ts 3Ts 4Ts 5Ts 6Ts

7Ts 8Ts 9Ts 10Ts

11Ts 12Ts 13Ts

Fig. 2.107 Output of the sample and hold circuit for sine input signal

t

132

2 Multirate Signal Processing

ho (t )

Fig. 2.108 Rectangle pulse signal

1 Ts

0 Fig. 2.109 Continuous time signal for sample and hold circuit

t

xc (t ) 16

8

8

0

16

20

t

Example 2.30 The signal shown in Fig. 2.109 is passed through a sample and hold circuit. Find the signal at the output of the sample and hold circuit. Take sampling period as Ts ¼ 2. Solution 2.30 First we determine the amplitude values for the time instants t such that t ¼ nTs where Ts ¼ 2 and n is integer. This operation result is shown in Fig. 2.110. In addition, we also write the line equations for the computation of the amplitude values for the given time instants. The amplitude values of the continuous time signal at time instants t ¼ nTs are shown clearly in Fig. 2.111. In the next step, we draw horizontal lines for the determined amplitudes, and for the first two samples, the drawn horizontal lines are shown in Fig. 2.112. And for the first 4 samples, the horizontal drawn lines are shown in Fig. 2.113. Repeating this procedure for all the other samples, we obtain the graph shown in Fig. 2.114. The drawn horizontal lines for all the samples are depicted alone in Fig. 2.115. Fig. 2.110 The continuous time signal in details

xc (t )

2t

16 14 12 10

t 24

2t 10

8

4

0

2 4 6

8 10 12 14 16 18 20

t

2.3 Practical Implementations of C/D and D/C … Fig. 2.111 Amplitudes shown explicitly for the time instants t ¼ nTs where Ts ¼ 2

133

xc (t ) 16 14 12 10 8

4

0

Fig. 2.112 Horizontal lines are drawn for the first two samples

2 4 6

8 10 12 14 16 18 20

2 4 6

8 10 12 14 16 18 20

2 4 6

8 10 12 14 16 18 20

2 4 6

8 10 12 14 16 18 20

t

xc (t ) 16 14 12 10

8

4

0

Fig. 2.113 Horizontal lines are drawn for the first four samples

t

xc (t ) 16 14 12 10

8 4

0 Fig. 2.114 Horizontal lines are drawn for all the samples

t

xc (t ) 16 14 12 10

8 4

0

t

134

2 Multirate Signal Processing

xo (t )

Fig. 2.115 Output of the sample and hold system

16 14 12 10 8

4

0

2.3.3

2 4 6

t

8 10 12 14 16 18 20

Quantization and Coding

During data storage or data transmission, we use bit sequences to represent real number. Since there are an infinite number of real numbers, it is not possible to represent this vast amount of real numbers by limited length bit streams. For this reason, we choose a number of real numbers to represent by bit streams and try to round other real numbers to the chosen ones when it is necessary to represent them by bit streams. Mid-Level Quantizer A typical quantizer includes the real number intervals used to map real numbers falling into these intervals to the quantization levels as shown in Fig. 2.116. The quantizer in Fig. 2.116 is called mid-level quantizer. The quantizer maps the real numbers in the range  D2 ; D2 to Q0 , maps the real numbers in the range D2 ; 3D 2 to Q1 etc. In this quantizer, D is called the step size of the quantizer. Smaller D means more sensitive quantizer. The mapping between real numbers and quantization levels is defined as Qi ¼ QðxÞ where Qi may be chosen as the center of interleaves. If Fig. 2.116 is inspected, it is seen that if we have equal number of intervals on the negative and positive regions, it means that the total number of intervals is an odd number, which is not a desired situation. Since using N bits, it is possible to represent 2N levels. For this reason, we design these quantizers such that if one side has even number of intervals, then the other side has odd number of intervals.

Q(x) Q

X m1

7 2

Q

3

5 2

Q0

Q1

2

3 2

Fig. 2.116 A typical mid-level quantizer

Q1

0

2

2

Q3

Q2

3 2

5 2

7 2

X m2

x

2.3 Practical Implementations of C/D and D/C …

135

Q (x) Q

Q4 9 2

Q

3

7 2

5 2

Q0

Q1

2

Q1

0

3 2

2

3 2

2

Q3

Q2

7 2

5 2

x

Fig. 2.117 Mid-level quantizer for Example 2.31

Q(x) Q

7 2

Q

3

5 2

Q1

2

Q0

Q1

0

3 2

2

2

Q2 3 2

Q3

5 2

Q4

7 2

x 9 2

Fig. 2.118 An alternative mid-level quantizer for Example 2.31

Example 2.31 A 3-bit quantizer includes 23 ¼ 8 quantization intervals. A mid-level type quantizer consisting of 8 levels can be shown as in Fig. 2.117. Or alternatively as in Fig. 2.118. We will use mid-level quantizers as in Fig. 2.117. As it is clear from the Example 2.30, for an N-bit mid-level quantizer, the minimum number that can be quantized is ð2N þ 1Þ=2 and the maximum number that can be quantized is ð2N  1Þ=2. The quantization levels are represented by binary sequences, such as two’s complement, one’s complements, unsigned representation, or private bit sequences can be assigned for quantization levels. Example 2.32 Design a 3-bit quantizer for the real numbers in the range ½14    14. Solution 2.32 For a 3-bit quantizer Xm1 ¼ 9D=2 and Xm2 ¼ 7D=2. Equating Xm2 to 14, we obtain 7D ¼ 14 ! D ¼ 4: 2 So our quantizer can quantize the real numbers in the range 

 9D 7D   ¼ ½18    14: 2 2

The bit sequences for our quantizer can be assigned to the intervals as in Fig. 2.119 and centers of the interleavers can be calculated as in Fig. 2.120. Mid-Rise Quantizer The mid-rise quantizer is shown in Fig. 2.121. As it is clear from Fig. 2.121, there is no interval centered at the origin.

136

2 Multirate Signal Processing

Q (x ) 001 Q3

000 Q 4 18

14

011 Q1

010 Q2

0

2

6

10

100 Q0

110 Q2

101 Q1 2

111 Q3

x 14

10

6

Fig. 2.119 Bit sequences assigned to the quantization intervals

Q (x ) 001 12

000 16 18

14

011 4

010 8

0

110

101 4

0

2

6

10

100 2

111 8

6

x 14

10

6

Fig. 2.120 Mid-level quantizer

Q (x ) Q

Q4

4

3

Q

3

2

Q1

Q1

0

2

Q4

Q3

Q2

3

2

x

4

Fig. 2.121 Mid-rise quantizer

Assume that we want to quantize a sequence of digital samples represented by x½n. Let ^x½n be the sequence obtained after quantization. Since quantization distorts the original signal, the quantized samples mathematically can be written as ^x½n ¼ Qðx½nÞ ! ^x½n ¼ x½n þ e½n

ð2:111Þ

where e½n is called quantization noise.

2.3.4

D/C Converter

The practical implementation of D/C converter is shown in Fig. 2.122. The content of the D/A converter is detailed in Fig. 2.123. The digital codes are converted to real numbers according to the used coding scheme. At the output of the code-to-digital converter, we have digital samples which can be written as

Fig. 2.122 D/C conversion Digital Code

Digital to Analog Conversion

D/ A

xo (t ) Reconstruction Filter

xr (t )

2.3 Practical Implementations of C/D and D/C …

137

Fig. 2.123 D/A conversion

Digital Code

^

x[n] x o (t)

Convert Digital Codes to Real Numbers

Digital Code

xo (t )

D/ A

(t nTs )

Zero Order Hold

xo (t )

Ts

n

ho (t )

Fig. 2.124 Impulse response of zero order hold

1

0

^x½n ¼ x½n þ e½n

Ts

t

ð2:112Þ

where e½n is the quantization error. The zero order hold filter impulse response is shown Fig. 2.124. The output of the code-to-digital converter in Fig. 2.123 is ^xo ðtÞ ¼

1 X

^x½ndðt  nTs Þ:

ð2:113Þ

n¼1

When ^xo ðtÞ is passed through zero order hold filter, we obtain xo ðtÞ ¼ ^xo ðtÞ * ho ðtÞ ! xo ðtÞ ¼

1 X

^x½nho ðt  nTs Þ:

ð2:114Þ

n¼1

Substituting ^x½n ¼ x½n þ e½n in (2.114), we get xo ð t Þ ¼

1 X n¼1

Fig. 2.125 Reconstruction filter block diagram

x½nho ðt  nTs Þ þ

1 X

e½nho ðt  nTs Þ:

ð2:115Þ

n¼1

x o (t )

Reconstruction Filter

xr (t )

138

2 Multirate Signal Processing

Now let’s consider the last unit of the D/C converter the reconstruction filter as shown in Fig. 2.125. The Fourier transform of xo ðtÞ in (2.115) can be calculated using Xo ðwÞ ¼

1 X

x½nHo ðwÞejwnTs þ

n¼1

1 X

e½nEo ðwÞejwnTs

ð2:116Þ

n¼1

where taking the common term Ho ðwÞ outside the parenthesis, we obtain 0

1

BX C 1 X B 1 C jwnTs jwnTs C x ½ n e þ e ½ n e Xo ðwÞ ¼ B B C @n¼1 A n¼1 |fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl} Xn ðTs wÞ

Ho ðwÞ

ð2:117Þ

En ðTs wÞ

which can be written as Xo ðwÞ ¼ ðXn ðTs wÞ þ En ðTs wÞÞHo ðwÞ:

ð2:118Þ

From Fig. 2.125, we can write Xr ðwÞ ¼ Hr ðwÞXo ðwÞ

ð2:119Þ

where Hr ðwÞ is the frequency response of the reconstruction filter. If we choose Hr ðwÞ as  Hr ðwÞ ¼

Ts H o ð wÞ

0

jwj\ Tps otherwise

ð2:120Þ

and substituting it into (2.119) and using (2.118) in (2.119), we obtain Xr ðwÞ ¼ Ts Xn ðTs wÞ þ Ts En ðTs wÞ jwj\

p Ts

ð2:121Þ

which is the Fourier transform of xr ðtÞ ¼ xa ðtÞ þ eðtÞ:

ð2:122Þ

Since x½n ¼ xa ðnTs Þ, e½n ¼ eðnTs Þ, the continuous time signals xa ðtÞ and eðtÞ can be obtained from their samples using

2.3 Practical Implementations of C/D and D/C …

139

  t  nTs xa ð t Þ ¼ x½n sin c Ts n¼1

ð2:123Þ

  t  nTs eð t Þ ¼ e½n sin c : Ts n¼1

ð2:124Þ

1 X

and 1 X

Then xr ðtÞ in (2.122) using (2.123) and (2.124) can be written as 1 X

xr ðtÞ ¼

n¼1

2.4

x½n sin c

    1 X t  nTs t  nTs e½n sin c þ : Ts Ts n¼1

Problems

(1) x½n ¼ ½1 2 0  3  1 1 4  1 0 1  2 5 1 3 is given. Find the signals x½2n, x½3n, x½4n, x½n=2, x½n=3, and x½n=4. (2) One period of the Fourier transform of x½n around origin is shown in Fig. 2.126. Draw the Fourier transform of the downsampled signal y½n ¼ x½2n. (3) The delay system is described in Fig. 2.127.

Fig. 2.126 One period of Xn ðwÞ around origin

X n (w)

w

1

3 4

Fig. 2.127 Delay system

x[n]

3 4

z

n0

w

x[n n0 ]

140

2 Multirate Signal Processing

Fig. 2.128 Signal processing system

x[n]

z

1

2

y1[n]

2

y2 [ n ]

If x½n ¼ ½a

b c

d

e

g

f

h

n¼0

ı j

k

l

m

n

o

p

r

find the output of each unit in Fig. 2.128. (4) Calculate the inverse Fourier transform of the digital filter  Hdn ðwÞ ¼

1 0

if if

wj\ Mp p M \jwj\p:

ð2:125Þ

(5) Draw the graph of  sin pn M hdn ½n ¼ pn

ð2:126Þ

roughly, and find the triangle approximation of (2.126). Calculate the approximated model for n ¼ 5; . . .; 5. (6) The graph of XðtÞ is shown in Fig. 2.129. Considering Fig. 2.129 draw the graph of

Y ðtÞ ¼

1 X

Xðt  kTÞ;

T ¼ 3:

ð2:127Þ

k¼1

X (t )

Fig. 2.129 The graph of XðtÞ

1

t

2

2

2.4 Problems

141

Fig. 2.130 Downsampler

fs

1000 Hz

4

y[n]

Fig. 2.131 System for Question 9 xc (t )

C/D

x[n]

fs 4

f ds

f ds

250 Hz

x[ Mn]

yr (t )

D/C

M

Ts

Ts

(7) Repeat Question-6 for T ¼ 1, T ¼ 4 and T ¼ 5. (8) Comment on the system shown in Fig. 2.130. (9) For the system of Fig. 2.131, xc ðtÞ is a lowpass signal with bandwidth 1 s and M ¼ 2. Is system output yr ðtÞ equal to system input 3000 Hz, Ts ¼ 8000 xc ðtÞ? If they are equal to each other, justify the reasoning behind it. If they are not equal to each other, again explain the reasoning behind it. (10) If x½n ¼ ½1 2 3 4 5 6 7 and L ¼ 4, draw the graph of

y ½ n ¼

1 X

x½kd½n  kL:

k¼1

(11) For the system of Fig. 2.132, M ¼ L ¼ 2 and x½n ¼ ½a

b c

d

e

f

g

h

l |{z}

j

k

l

m

n

o

p

r

s:

n¼0

Find xd ½n and y½n. (12) Draw the graph of hai ½n ¼  jLnj þ 1, L  n  L for L ¼ 3 and L ¼ 8. (13) xd ½k ¼ ½1 and draw

4

7 10 13, hai ½n ¼  jLnj þ 1, L  n  L, L ¼ 3, calculate

Fig. 2.132 Signal processing system

xc (t )

C/D

Ts

x[n] M

xd [n]

y[n] L

142

2 Multirate Signal Processing 1 X

yi ½n ¼

xd ½khai ½n  kL:

k¼1

 (14) For the system of Fig. 2.133, x½n ¼ cos p4 n 0  n  10, hai ½n is the triangle approximated reconstruction filter. Find xd ½n; y½n and yi ½n for M ¼ L ¼ 2. (15) For the system of Fig. 2.134,  Haa ðwÞ ¼

if jwj\ Tps otherwise

1 0

Express the Fourier transform of x½n in terms of the Fourier transform of xc ðtÞ. (16) For the system of Fig. 2.135, M ¼ 3, Xn ðwÞ is the one period of the Fourier transform of x½n. Draw the Fourier transform of xd ½n. (17) For the system of Fig. 2.136, M ¼ L ¼ 2 and Xn ðwÞ is the one period of the Fourier transform of x½n.

x[n]

xd [n]

M

Downsampler Compressor

y[n] L Upsampler Expander

hai [n ]

yi [n]

Interpolating Filter

Fig. 2.133 Signal processing system for Question 14

xc (t )

H aa (w)

C/D

x[n]

Ts Fig. 2.134 Signal processing system for Question 15

X n (w) 1

x[n]

3

2 3

Fig. 2.135 Downsampling of digital signal

w

M Downsampling

xd [n]

2.4 Problems

143 X n ( w) 1

2

x[n]

hd [n]

xc [n]

Decimator Filter

w

2 3 xd [n]

M

Downsampling

y[n]

L

hi [n]

yi [n]

Interpolation Filter

Upsampling

Fig. 2.136 Signal processing system for Question 17

x[n]

xc [n]

hd [n] Decimator Filtre

M

xd [n]

Downsampling

Fig. 2.137 Decimation system

(a) (b) (c) (d)

Draw the Fourier transforms of xc ½n; xd ½n, and y½n. Draw the triangle approximation model of the interpolation filter for L ¼ 2. Draw the Fourier transform of yi ½n for sin cðÞ interpolation filter. If xc ½n ¼ ½1:0 1:7 2:4 3:2 4, calculate xd ½n; y½n, using triangle approximated interpolation filter.

(18) For the system of Fig. 2.137, M ¼ 2, and Hd ðwÞ is defined as  H d ðw Þ ¼

1 0

if jwj  otherwise

p M

ð2:128Þ

(a) Calculate the inverse Fourier transform of Hd ðwÞ, i.e., calculate hd ½n. Next, find the triangle approximated model of hd ½n. (b) For x½n ¼ ½ 1 2 3 4  calculate xo ½n and xd ½n.

x[n] Fig. 2.138 Delay system

z

n0

x[n n0 ]

144

2 Multirate Signal Processing

xa [n]

x[n]

M

xc [n]

M

xe [n]

xr [n]

xg [n] z

z

xb [n]

M

xd [n]

M

1

x f [n] x j [n]

z

z xg [n]

M

xh [n]

M

1

xi [n]

Fig. 2.139 Signal processing system for Question 19

(19) The delay system is shown in Fig. 2.138. For the system of Fig. 2.139, M ¼ 3, x½n ¼ ½a b c d e f g h i j k l m n o p r s t u v w x y: Find the signal at the output of each unit, and find the system output.

Chapter 3

Discrete Fourier Transform

In linear algebra, basis vectors span the entire vector space. And any vector of the vector space can be written as the linear combination of the basis vectors. For any vector in vector space, finding the coefficients of basis vectors used for the construction of the vector can be considered as a transformation. Fourier series are used to represent periodic signals. Fourier series are used to construct any periodic signal from sinusoidal signals. The sinusoidal signals can be considered as the basis signals, and linear combination of these signals with complex coefficients produce any periodic signal. Once we obtain the coefficients of the base signals necessary for the construction of a periodic signal, then we have full knowledge of the periodic signal and instead of transmitting the periodic signal, we can transmit the coefficients of the base signals. Since at the receiver side, the periodic signal can be reconstructed using the base coefficients. In this chapter we will study a new transformation technique called discrete Fourier transform used for aperiodic digital signals. We will show that similar to the Fourier series representation of periodic digital signals, aperiodic digital signals can also be written as a linear combination of sinusoidal digital aperiodic signals. In this case, aperiodic digital sinusoidal signals can be considered as base signals. And finding the coefficients of base signals such that their linear combination yields the aperiodic digital signal is called discrete Fourier transformation of the aperiodic digital signal. Thus, the discrete Fourier transformation is nothing but finding the set of coefficients of the base signals for an aperiodic digital signal. And once we have these coefficients, then we have full knowledge on the aperiodic signal in another digital sequence.

© Springer Nature Singapore Pte Ltd. 2018 O. Gazi, Understanding Digital Signal Processing, Springer Topics in Signal Processing 13, DOI 10.1007/978-981-10-4962-0_3

145

146

3

3.1

Discrete Fourier Transform

Manipulation of Digital Signals

Before studying discrete Fourier transform, let’s prepare ourselves for the subject, for this purpose, we will first study the manipulation of digital signals. Manipulation of Non-periodic Digital Signals A non-periodic or aperiodic digital signal has finite number of samples. And these signals are illustrated either by graphics or by number vectors, or by number sequences. As an example, a digital signal and its vector representation is shown in Fig. 3.1. Manipulation of digital signals includes shifting, scaling in time domain and change in amplitudes. Shifting of Digital Signals in Time Domain Given x½n, to obtain x½n  n0 ; n0 [ 0, we shift the amplitudes of x½n to the right by n0 units. If n0 \0, amplitudes are shifted to the left. Shifting amplitudes to the right by n0 equals to the shifting n ¼ 0 index to the left by n0 units. This operation is illustrated in the following example. Example 3.1 Given x½n ¼ ½a b c d x½n  3;

x½n þ 1;

e |{z}

f

g

j k; find x½n  1

h i

n¼0

x½n þ 2; and x½n  7:

Solution 3.1 To get x½n  1, we shift amplitudes of x½n to the right by ‘1’ unit. Shifting amplitudes to the right by ‘1’ unit is the same as shifting n ¼ 0 index to the left by ‘1’ unit, the result of this operation is x½n  1 ¼ ½a

b

c

d |{z}

e

f

g

h i

j

k:

j

k:

n¼0

Following a similar approach for x½n  3; we obtain x½n  3 ¼ ½a

c

b |{z}

d

e

f

g

h

i

n¼0

x[n]

1.7

3

3

2

1.5 2

1

0

1

2

2.5 1.25 3

4

x[n] [1.7 1.5

3

n

1.5

3

Fig. 3.1 A digital signal and its representation by a number vector

3 n 0

1.5

2

2.5

1.25]

3.1 Manipulation of Digital Signals

147

To get x½n þ 1, we shift amplitudes of x½n to the left by ‘1’ unit obtaining x½n þ 1 ¼ ½a b

c

d

e

f |{z}

g

h

i

j

k:

g |{z}

h

i

j

k;

n¼0

Following similar steps, we obtain x½n þ 2 ¼ ½a

b

c d

e

f

n¼0

and x½n  7 ¼ ½|{z} 0

0

0 a

b

c

d

e

f

g

h

i

j

k:

n¼0

where it is clear that if the shifting amount goes beyond the signal frontiers, for the new time instants, 0 values are assigned for the signal amplitudes. Scaling of Digital Signals in Time Domain To find x½Mn, we divide the time axis of x½n by M, and keep only integer division results and omit the non-integer division results. The resulting signal is nothing but x½Mn. Example 3.2 If x½n ¼ ½a b

c

d

e |{z}

f

g

h

i

k; find x½2n and

j

n¼0

x½3n. Solution 3.2 To get x½2n; we divide time axis of x½n by 2 and keep only integer division results. First, let’s write all the time indices as shown in ½|{z} a 4

b |{z} 3

c |{z} 2

d |{z} 1

e |{z} n¼0

f |{z}

g |{z}

1

2

h |{z}

i |{z}

3

4

j |{z} 5

k : |{z} 6

ð3:1Þ Next, we divide the indices as in ½|{z} a

b |{z}

c |{z}

d |{z}

e |{z}

42

32

22

12

0 2

f |{z}

g |{z}

1 2

2 2

h |{z}

i |{z}

3 2

4 2

j |{z} 5 2

k  |{z} 6 2

ð3:2Þ where keeping only integer division results, we obtain x½2n ¼ ½|{z} a 2

b |{z} 2

c |{z} 0

g |{z} 1

i |{z}

k  |{z}

2

3

148

3

Discrete Fourier Transform

which can be written in its simple form as x½2n ¼ ½a

c |{z} e

g

i

k:

0

Following a similar approach for x½3n, we obtain x½3n ¼ ½b

h

e |{z}

k:

n¼0

Combined Shifting and Scaling To obtain x½Mn  n0 , we follow a two-step procedure as listed below. (1) First, the shifted signal, x½n  n0  is obtained, and this signal is denoted by x1 ½n; i.e., x1 ½n ¼ x½n  n0  (2) Then using x1 ½n, we obtain the scaled signal x1 ½Mn which is nothing but x½Mn  n0  That is, we first obtain the shifted signal x1 ½n ¼ x½n  n0 , and then using x1 ½n we get the scaled signal x1 ½Mn ¼ x½Mn  n0 : Example 3.3 If x½n ¼ ½a

b

c d

e |{z}

f

g h

j k, find x½3n þ 3.

i

n¼0

Solution 3.3 First, we obtain the shifted signal x½n þ 3 as x½n þ 3 ¼ ½a

b

c d

e

f

g

h |{z}

i

j k:

n¼0

Let x1 ½n ¼ x½n þ 3; i.e., x1 ½n ¼ ½a

b

c

d

e f

e

h |{z}

h |{z}

i

j k;

n¼0

then the scaled signal x1 ½3n can be calculated as x1 ½3n ¼ ½b

g

k

n¼0

which is nothing but x½3n þ 3; that is x½3n þ 3 ¼ ½b e

h |{z}

k:

n¼0

Note: If n ¼ 0 index is not indicated in the digital signal vector representation, then the first element index is accepted as n ¼ 0:

3.1 Manipulation of Digital Signals

3.1.1

149

Manipulation of Periodic Digital Signals

Manipulation of periodic digital signals includes shifting, scaling and combined shifting, scaling operations. There is no difference in manipulating non-periodic and periodic digital signals. The same set of operations are applied for the manipulation of periodic signals as in the manipulation of non-periodic signals. However, since periodic signals are of infinite lengths, for easy of manipulation, it is logical to consider just one period of the periodic signal and perform manipulations on it. Let ~x½n be a periodic signal with fundamental period N; i.e., ~x½n ¼ ~x½n þ lN l; N 2 Z. Let’s define one period of ~x½n as  x ½ n ¼

~x½n 0

0nN  1 otherwise:

ð3:3Þ

Using (3.3), we can write ~x½n in terms of x½n as 1 X

~x½n ¼

x½n  kN:

ð3:4Þ

k¼1

3.1.2

Shifting of Periodic Digital Signals

First let’s make definitions as follows: Rotate Right When the signal x½n ¼ ½ 1 2 3 4 RRðx½nÞ ¼ ½ N

1

2

   N  is rotated right, we get 3 4

   N  1 :

ð3:5Þ

RRðx½n; mÞ is the m unit rotated (right) signal. Rotate Left When the signal x½n ¼ ½ 1 2 3 4    N  is rotated left, we get RLðx½nÞ ¼ ½ 2 3

4

 N  1

N

1 :

RLðx½n; mÞ is the m unit rotated (left) signal. Rotate Inside When the signal x½n ¼ ½ 1 2 3 4    N  is rotated inside, we get

ð3:6Þ

150

3

RI ðx½nÞ ¼ ½ 1

N 1

N

N2

Discrete Fourier Transform

   2 :

ð3:7Þ

Shifting of Periodic Digital Signals If x½n is the one period of the periodic signal, ~x½n such that 0  n  N  1; one period of the shifted signal ~x½n  n0 , n0 [ 0 is obtained by rotating amplitudes of x½n to the right (left if n0 \0) by n0 units. Example 3.4 The signal given in Fig. 3.2 is a periodic signal, i.e., ~x½n ¼ ~x½n þ N. Find the period of this signal, and determine its one period for 0  n  N  1. Solution 3.4 To find the period of the signal, we need to find the repeating pattern in the signal graph. If the signal shown in Fig. 3.2 is carefully inspected the repeating pattern can be easily determined. The repeating pattern of Fig. 3.2 is shown in Fig. 3.3 in bold. The number of samples in the repeating pattern is nothing but the period of the signal. Hence, for this example, N the period of the signal is 5, i.e., ~x½n ¼ ~x½n þ 5: One period of the signal in Fig. 3.3 for 0  n  4 is shown in Fig. 3.4. Using one period of the signal starting at origin, we can write the periodic signal as ~x½n ¼ ½  

3 |{z}

1:5

1:7

1:5

3

  :

n¼0

x[n]

Fig. 3.2 A periodic digital signal

3

1.7

3

1.7

1.5

2

1

0

1 2

3

1.5 3

4 5

n

1.5

1.5

3.5

3.5

Fig. 3.3 The repeating pattern of Fig. 3.2 is shown in bold

6

x[n] 3

1 .7

3

1 .7

1 .5 2

1

0

1 2

3

1 .5 3

4 5

1 .5

1 .5

3 .5

6

3 .5

n

3.1 Manipulation of Digital Signals Fig. 3.4 One period of the signal in Fig. 3.3 for 0n4

151

x[n] 3 1 .7

3

1 .7

1 .5 2

0

1

3

1.5

4 5

3

1 2

1 .5

1 .5

3.5

3. 5

Fig. 3.5 The periodic signal ~x½n for Example 3.5

n

6

x[n] 3

1.7

3

1.7

1 .5

2

0

1

3 1 .5

4 5

3

1 2

n

6

1. 5

1.5

3.5

3.5

Example 3.5 The periodic signal ~x½n is shown in Fig. 3.5, find ~x½n  3; and ~x½n þ 2. Solution 3.5 The period of the signal is N ¼ 5, and signal amplitudes for one period are x½n ¼ ½|{z} 3

1:5

1:7

3:5:

1:5

ð3:8Þ

n¼0

When x½n is rotated to the right by 3 units, we get RRðx½n; 3Þ ¼ ½|{z} 1:7

1:5

3:5

3

1:5:

ð3:9Þ

n¼0

And using (3.9), we can write the shifted periodic signal as ~x½n  3 ¼ ½  

1:7 1:5

3:5

3

1:5

1:7 |{z}

1:5

3:5 3

1:5

  :

n¼0

ð3:10Þ To find ~x½n þ 2, one period of ~x½n is rotated to the left by 2 units yielding RLðx½n; 2Þ ¼ ½|{z} 1:7 n¼0

1:5

3:5

3

1:5:

ð3:11Þ

152

3

Discrete Fourier Transform

Hence, shifted periodic signal ~x½n þ 2 becomes as ~x½n þ 2 ¼ ½   1:5

3:5

3

1:5

1:7 |{z}

1:5

3:5 3

1:5

1:7   :

n¼0

ð3:12Þ Time Scaling of Periodic Signals To perform time scaling on periodic signals, we consider one period of the signal and perform time scaling on it. The resulting signal is nothing but the one period of the scaled signal. If the period of the digital signal ~x½n is N, then the period of the scaled signal ~x½Mn is N=M. Example 3.6 The periodic signal ~x½n in its one interval equals to x½n ¼ ½|{z} 3

1:5

1:7 1:5

3:5

2:2

ð3:13Þ

4

n¼0

where it is obvious that the period of the signal is N ¼ 7. Find ~x½2n and ~x½3n. Solution 3.6 One period of ~x½2n equals to x½2n, and one period of ~x½3n equals to x½3n. The time scaled signals x½2n and x½3n can be calculated as x½2n ¼ ½|{z} 3

1:7 3:5 4

n¼0

x½3n ¼ ½|{z} 3

ð3:14Þ

1:5 4:

n¼0

And using (3.14) the periodic signals ~x½2n and ~x½3n can be written as ~x½2n ¼ ½   1:7 3:5

4 |{z} 3

1:7 3:5

4 3

1:7

3:5

4   

n¼0

~x½3n ¼ ½   3

1:5 4

3 |{z}

1:5

4

3

1:5

4   :

n¼0

Combined Shifting and Scaling The periodic digital signal ~x½n can be shifted and scaled in time domain yielding the periodic signal ~x½Mn  n0 . The shifted and scaled signal ~x½Mn  n0  can be obtained from ~x½n via a two-step procedure as explained below. (1) To get ~x½Mn  n0 ; first the shifted signal ~x½n  n0  is obtained. Let’s call this signal ~x1 ½n, i.e., ~x1 ½n ¼ ~x½n  n0 . (2) In the next step, ~x1 ½n is scaled in time domain and ~y½n ¼ ~x1 ½Mn is obtained, and ~y½n is nothing but ~x½Mn  n0 , i.e., ~y½n ¼ ~x½Mn  n0 .

3.1 Manipulation of Digital Signals

153

Example 3.7 The periodic signal ~x½n in its one interval equals to x½n ¼ ½|{z} 3

1:5

1:7 1:5

3:5

2:2

4

n¼0

where it is obvious that the period of the signal is N ¼ 7. Find ~x½2n  3. Solution 3.7 To obtain ~x½2n  3, let’s first find one period of the shifted signal ~x½n  3: One period of ~x½n  3 is obtained by rotating one period of ~x½n to the right by 3 yielding RRðx½n; 3Þ ¼ ½3:5 |ffl{zffl}

2:2 4

3

1:5

1:7

ð3:15Þ

1:5

n¼0

Let’s denote (3.15) by x1 ½n, i.e., one period of ~x½n ¼ ~x½n  3, then we have x1 ½n ¼ ½3:5 |ffl{zffl}

2:2

3 1:5

4

1:7

ð3:16Þ

1:5:

n¼0

Next using (3.16), we can evaluate x1 ½2n which is nothing but one period of ~x½2n  3 as x1 ½2n ¼ ½3:5 |ffl{zffl}

4

1:5

1:5:

n¼0

Hence, our periodic signal ~x½2n  3 becomes as ~x½2n  3 ¼ ½    1:5

1:5 |ffl{zffl} 3:5

1:5

4

1:5

3:5

1:5   :

4

n¼0

Example 3.8 Periodic signal ~x½n is shown in Fig. 3.6. Find ~x½n. Solution 3.8 To find ~x½n, we divide the time axis of ~x½n by 1. This operation is illustrated in Fig. 3.7. The division operations in Fig. 3.7 yields the signal in Fig. 3.8. When amplitudes and time indices are re-ordered together, we obtain the graph in Fig. 3.9. Practical way to find ~x½n signal Fig. 3.6 Periodic signal ~x½n for Example 3.8

x[n]

a

b

c

d

a

b

c

d

a

4

3

2

1

0

1

2

3

4

n

154

3

Fig. 3.7 Calculation of ~x½n

Fig. 3.8 After division of the time axis in Fig. 3.7

Fig. 3.9 Time axis re-ordered

Fig. 3.10 Periodic signal ~x½n

a

b

4 1

3 1

a

b

c

d

4

3

2

1

c

x[-n] a b

d

2 1

Discrete Fourier Transform

1 1

c

0 1 2 1 1 1

d

a

x[-n] a b

c

d

a

1

2

3

4

c

b

a

1

2

3

4

0

a

d

c

x[-n] b a d

4

3

2

1

a

b

c

d

a

b

c

d

a

4

3

2

1

0

1

2

3

4

0

x[n]

n

3 4 1 1

n

n

n

If one period of ~x½n is denoted by x½n ¼ ½ 1 2 3 4    N ; then one period of ~x½n can be obtained rotating x½n inside by 1 unit. That is, one period of ~x½n is RI ðx½nÞ ¼ ½ 1 N

N1

N2

 2

ð3:17Þ

We can apply this practical method to the previous example where the periodic signal had been given as in Fig. 3.10. One period of is ~x½n in Fig. 3.10 is x½n ¼ ½ a

b

c

d :

ð3:18Þ

When (3.18) is rotated inside, we obtain RRðx½nÞ ¼ ½ a

b

d

c

ð3:19Þ

which is nothing but one period of ~x½n. Hence ~x½n equals to ~x½n ¼ ½   d

c

b

a |{z}

d

c

n¼0

Calculation of the periodic signal ~x½n0  n

b

a

d

c b   :

3.1 Manipulation of Digital Signals Fig. 3.11 The periodic signal ~x½n for Example 3.9

155 x[n]

a

b

c

d

a

b

c

d

a

4

3

2

1

0

1

2

3

4

n

Calculation of the periodic signal ~x½n0  n can be achieved via the following steps. (1) We first find one period of ~x1 ½n ¼ ~x½n using rotate inside operation. (2) Then one period of ~x1 ½n is rotated to the right if n0 [ 0 to the left if n0 \0 by jn0 j units and one period of ~x1 ½n0  n is obtained. Example 3.9 The periodic signal ~x1 ½n is shown in Fig. 3.11. Find ~x1 ½2  n. Solution 3.9 From Fig. 3.11 one period of ~x½n can be found as x½n ¼ ½ a

b

d :

c

ð3:20Þ

When (3.20) is rotated inside, we obtain RRðx½nÞ ¼ ½ a

d

c

b

ð3:21Þ

which is nothing but one period of ~x½n, i.e., ~x½nop ¼ ½ a d c b , ‘op’ means one period. To find one period of ~x½2  n one period of ~x½n is rotated to the right by 2 units yielding   RR ~x½nop ; 2 ¼ ½ a b

c

d :

ð3:22Þ

Using (3.22) the periodic signal ~x½2  n can be written as ~x½2  n ¼ ½   c

b

a d

b

c |{z}

a d

c

b

a

d   :

n¼0

Example 3.10 The periodic signal ~x½n is shown in Fig. 3.11. Find ~x½2  n. Fig. 3.12 Solution 3.10 One period of ~x½n equals to x½n ¼ ½ a

Fig. 3.12 The periodic signal ~x½n for Example 3.10

b

d :

c

ð3:23Þ

x[n]

a

b

c

d

a

b

c

d

a

4

3

2

1

0

1

2

3

4

n

156

3

Discrete Fourier Transform

When (3.23) is rotated inside, we obtain RRðx½nÞ ¼ ½ a

d

b

c

which is nothing but one period of ~x½n, i.e., ~x½nop ¼ ½ a d c b : To find one period of ~x½2  n, one period of ~x  n is rotated to the left by 2 units yielding   RR ~x½nop ; 2 ¼ ½ c b

a

d :

ð3:24Þ

Using (3.24) the periodic signal ~x½2  n can be written as ~x½2  n ¼ ½   c

b

a d

b

c |{z}

a d

c

b

a d   :

n¼0

Exercise: For the previous exercise find ~x½4  n and ~x½4  n.

3.1.3

Some Well Known Digital Signals

In this subsection, we will review some well-known digital signals. Unit Step: The unit step signal is defined as  u½ n ¼

1 0

if n  0 otherwise

ð3:25Þ

if n ¼ 0 otherwise

ð3:26Þ

whose graph is shown in Fig. 3.13. Unit Impulse: The unit impulse signal is defined as  d½ n ¼

1 0

whose graph is shown in Fig. 3.14. u[n]

Fig. 3.13 Unit step function, i.e., signal

3

2

1

1

1

1

1

0

1

2

3

n

3.1 Manipulation of Digital Signals

157 [n]

Fig. 3.14 Unit impulse function, i.e., signal 2

3

1

1

0

2

1

3

n

The relation between u½n and d½n can be written as d½n ¼ u½n  u½n  1

ð3:27Þ

or as u½ n ¼

1 X

d½n  k

ð3:28Þ

k¼0

which is equal to u½ n ¼

n X

ð3:29Þ

d½k:

k¼1

Exponential Digital Signal The exponential digital signal is defined as ð3:30Þ

x½n ¼ ejw0 n which can also be written in the form x½n ¼ cosðw0 nÞ þ j sinðw0 nÞ:

ð3:31Þ

Example 3.11 Simplify ejk2p . Solution 3.11 Using (3.31), we have ejk2p ¼ cosðk2pÞ þ j sinðk2pÞ ¼ cosðk2pÞ þ j sinðk2pÞ |fflfflfflfflffl{zfflfflfflfflffl} |fflfflfflfflfflffl{zfflfflfflfflfflffl} ¼1

As a special case for k ¼ 1, we have e

j2p

¼0

¼ 1:

Example 3.12 Verify the following equality N 1 X k¼0

ej N km ¼ 2p



N 0

if m ¼ 0 otherwise:

ð3:32Þ

158

3

Discrete Fourier Transform

Solution 3.12 Let’s open the summation expression in (3.32) as follows   N 1 X j2pðN1Þm N j2p km j2p m j2p 2m N N N e ¼ 1þe þe þ  þe :

ð3:33Þ

k¼0

The right hand side of (3.33) can be simplified using the property 1 þ x þ x2 þ x3 þ    þ xN1 ¼

1  xN 1x

ð3:34Þ

as in 

1þe

j2p Nm

þe

j2p N 2m

þ  þe

j2p N ðN1Þm



1  ej N mN 2p

¼

1e

j2p Nm

¼

1  ej2pm 1  ej N m 2p

:

ð3:35Þ

And for m 6¼ 0 using the result in, (3.35), we obtain 1  ej2pm 1  ej N m 2p

¼

11 1  ej N m 2p

! 0:

ð3:36Þ

Hence we have N 1 X

ej N km ¼ 0; 2p

m 6¼ 0:

ð3:37Þ

k¼0

And for m ¼ 0 using the result in (3.35), we obtain N 1 X

ej N km ¼ 2p

k¼0

N 1 X

1 ! N:

ð3:38Þ

if m ¼ 0 otherwise:

ð3:39Þ

k¼0

Combining (3.37) and (3.38), we obtain N 1 X

e

j2p N km

k¼0

3.2

 ¼

N 0

Review of Signal Types

Basically we can divide signals into two categories as, continuous and digital signals. And in both classes, we can have periodic and non-periodic (aperiodic) signals, and Fourier transform and representation methods are defined for these classes of signals. In Fig. 3.15; the relation between signals and their transform or representation types are summarized.

3.2 Review of Signal Types

159 Signals

Continuous Time Signals

Digital Signals

Periodic Continuous Time Signals

Aperiodic Continuous Time Signals

Fourier Series Representation

Fourier Transform

Periodic Digital Signals

Aperiodic Digital Signals

Discrete Time Fourier Series Representation

Discrete Time Fourier Transform

Discrete Time Fourier Transform

Discrete Fourier Trasform

Fourier Transform

Fig. 3.15 Signals types, their transformations and representations

Let’s briefly review the signal types, their transformations and representations. Non-periodic Continuous Time Signals If xc ðtÞ is a non-periodic continuous time signal, then its Fourier is defined as Z1 X c ðw Þ ¼

xc ðtÞejwt dt

ð3:40Þ

1

and its inverse Fourier transform is given as 1 xc ð t Þ ¼ 2p

Z1 Xc ðwÞejwt dw

ð3:41Þ

1

where w ¼ 2pf is the angular frequency. The Fourier transform and inverse Fourier transform pairs show small differences in their coefficients in literature. In general, Fourier transform and inverse Fourier transform can be defined as Z1 Xc ðwÞ ¼ K1 1

xc ðtÞejwt dt

ð3:42Þ

160

3

Discrete Fourier Transform

and Z1 xc ð t Þ ¼ K 2

Xc ðwÞejwt dw

ð3:43Þ

1

where 1 : ð3:44Þ 2p pffiffiffiffiffiffi pffiffiffiffiffiffi Thus if K1 ¼ 1= 2p, then K2 should be 1= 2p so that K1  K2 ¼ 1=2p. As another example if K1 ¼ 1=2p then K2 ¼ 1: Periodic Continuous Time Signals If ~xc ðtÞ is a periodic signal with fundamental period T, then K1  K2 ¼

~xc ðtÞ ¼ ~xc ðt þ mT Þ:

ð3:45Þ

And for the periodic signal ~xc ðtÞ the Fourier series representation is defined as ~xc ðtÞ ¼

1 1 X 2p ~x½k ejk T t T k¼1

where the Fourier series coefficients ~x½k  are computed by using Z 2p ~xc ½k ¼ ~xc ðtÞejk T t dt:

ð3:46Þ

ð3:47Þ

T

If we define 2p=T by w0 , i.e., w0 ¼ 2p=T, then the above equations can also be written as ~xc ðtÞ ¼

1 1 X ~xc ½k ejkw0 t T k¼1

ð3:48Þ

and Z ~xc ½k  ¼

~xc ðtÞejkw0 t dt

ð3:49Þ

T

In general, the Fourier series representation of ~xc ðtÞ and its Fourier series coefficients are given as

3.2 Review of Signal Types

161 1 X

~xc ðtÞ ¼ K1

~xc ½k ejkw0 t

ð3:50Þ

~xc ðtÞejkw0 t dt

ð3:51Þ

k¼1

and Z ~xc ½k ¼ K2 T

pffiffiffiffi where the coefficients satisfy K1  K2 ¼ 1=T. Hence, if K1 ¼ 1= T then K2 ¼ pffiffiffiffi 1= T and Fourier series representation and Fourier coefficients expressions becomes as 1 1 X ~xc ðtÞ ¼ pffiffiffiffi ~xc ½kejkw0 t T k¼1

ð3:52Þ

and 1 ~xc ½k ¼ pffiffiffiffi T

Z

~xc ðtÞejkw0 t dt:

ð3:53Þ

T

Now let’s assume that one period of ~xc ðtÞ is xc ðtÞ, i.e., xc ðtÞ is an aperiodic signal. Then the Fourier series coefficients of ~xc ðtÞ is computed as Z ~xc ½k ¼

~xc ðtÞe

jkw0 t

Z1 dt ! ~xc ½k ¼

xc ðtÞejkw0 t dt:

ð3:54Þ

1

T

And the Fourier transform of xc ðtÞ is Z1 X c ðw Þ ¼

xc ðtÞejwt dt

ð3:55Þ

1

When (3.54) and (3.55) are compared to each other as in Z1 ~xc ½k ¼ 1

xc ðtÞejkw0 t dt $ Xc ðwÞ ¼

Z1 1

xc ðtÞejwt dt

ð3:56Þ

162

3

Discrete Fourier Transform

we see that ~xc ½k ¼ Xc ðwÞjw¼kw0

ð3:57Þ

where w0 ¼

2p : T

ð3:58Þ

And the relation between ~xc ðtÞ and xc ðtÞ can be written as 1 X

~xc ðtÞ ¼

xc ðt  kT Þ:

ð3:59Þ

k¼1

The Fourier transform of the periodic continuous time signal is defined as ~xc ðwÞ ¼

1 2p X ~xc ½kdðw  kw0 Þ; T k¼1

w0 ¼ 2p=T:

ð3:60Þ

Aperiodic Digital Signals The discrete time Fourier transform for the aperiodic digital signal x½n is defined as Xn ðwÞ ¼

1 X

x½nejwn

ð3:61Þ

n¼1

where w ¼ 2pf is the angular frequency, and the inverse Fourier transform is defined as 1 x½n ¼ 2p

Z Xn ðwÞejwn dw:

ð3:62Þ

2p

The Fourier transform function of x½n, i.e., Xn ðwÞ is a continuous function of w and it is also a periodic function with period 2p, i.e., Xn ðwÞ ¼ Xn ðw þ k2pÞ:

ð3:63Þ

Periodic Digital Signals If the digital signal ~x½n is a periodic signal, then ~x½n ¼ ~x½n þ lN  l; N 2 Z and N is called fundamental period of ~x½n: For the digital periodic signal ~x½n, the Fourier series representation is defined as

3.2 Review of Signal Types

163

~x½n ¼

1X 2p ~xn ½kejk N n N k;N

ð3:64Þ

where the Fourier series coefficients are computed using ~xn ½k ¼

X

~x½nejk N n : 2p

ð3:65Þ

n;N

Note:

P

ðÞ means summation is taken over any interval of length N; i.e.,

n;N

summation is taken over one period length. In general, the Fourier series representation and calculation of Fourier series coefficient of periodic signals are done via ~x½n ¼ K1

X

2p

~xn ½k ejk N n

ð3:66Þ

~x½nejk N n

ð3:67Þ

k;N

and ~xn ½k  ¼ K2

X

2p

n;N

such that K1  K2 ¼

1 : N

ð3:68Þ

The Fourier transform of the periodic digital signal ~x½n is ~xðwÞ ¼

1 2p X ~xn ½kdðw  kw0 Þ; N k¼1

w0 ¼

2p : N

ð3:69Þ

Example 3.13 If the Fourier series representation of digital periodic signal ~x½n is ~x½n ¼

1X 2p ~xn ½kejk N n N k;N

ð3:70Þ

then verify that the Fourier series coefficients as obtained using ~xn ½k ¼

X n;N

~x½nejk N n : 2p

ð3:71Þ

164

3

Discrete Fourier Transform

Solution 3.13 If the Fourier series coefficients are obtained using X

~x½nejk N n

ð3:72Þ

1X 2p ~xn ½kejk N n N k;N

ð3:73Þ

~xn ½k ¼

2p

n;N

then when (3.72) is substituted into ~x½n ¼

we should get ~x½n on the right hand side of (3.73). That is ~x½n ¼

1 XX 2p 2p ~x½r ejk N r ejk N n N k;N r;N |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} ~xn ½k

¼ ¼ ¼

N 1 X N 1 1X

N 1 N

2p

2p

k¼0 r¼0 N1 X N1 X

~x½r ejk N ðrnÞ 2p

ð3:74Þ

k¼0 r¼0

N 1 1X

N

~x½r ejk N r ejk N n

N 1 X

~x½r 

r¼0

ejk N ðrnÞ 2p

k¼0

¼

|fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl}  N if r ¼ n 0

otherwise

1 ¼ N~x½n N ¼ ~x½n Convolution of Aperiodic Digital Signals For aperiodic digital signals x½n; y½n, the convolution operation is defined as x½n y½n ¼

1 X

x½k y½n  k

ð3:75Þ

x½n  k y½k:

ð3:76Þ

k¼1

or x ½ n y ½ n ¼

1 X k¼1

3.3 Convolution of Periodic Digital Signals

3.3

165

Convolution of Periodic Digital Signals

Let ~xn ½n and ~x2 ½n be digital periodic signals with common period N, i.e., ~x1 ½n ¼ ~x1 ½n þ N  and ~x2 ½n ¼ ~x2 ½n þ N : The period convolution of ~x1 ½n and ~x2 ½n is defined as N 1 X

~x3 ½n ¼

~x1 ½m~x2 ½n  m:

ð3:77Þ

m¼0

The digital sequence ~x3 ½n is also periodic with period N. How to calculate periodic convolution? This is explained as follows. (1) Since ~x3 ½n is periodic with the same period N; we can focus on the calculation of one period of ~x3 ½n starting from 0, i.e., consider 0  n  N  1. (2) When the summation in (3.77) is expanded, we get ~x3 ½n ¼ ~x1 ½0~x2 ½n þ ~x1 ½1~x2 ½n  1 þ    þ ~x1 ½N  1~x2 ½n  ðN  1Þ

ð3:78Þ

where we can use only one period of ~x2 ½n; ~x2 ½n  1; and ~x2 ½N  1; 0  n  N  1. Example 3.14 The periodic signals ~x1 ½n and ~x2 ½n with period N ¼ 4 are shown in Fig. 3.16. Calculate their 4-point periodic convolution. Solution 3.14 The periodic convolution for the given signals is calculated using N 1 X

~x3 ½n ¼

~x1 ½m~x2 ½n  m:

ð3:79Þ

m¼0

~

x1[n]

2 1

1

4

3

2

1

2

1

1

0

1

1

2

3

4

n

1

1 ~

x2 [n]

4 1

1

1

3

2

1 1

0

1

1

1

2

3

1

Fig. 3.16 The periodic signals ~x1 ½n and ~x2 ½n for Example 3.14

1

4 1

n

166

3

Discrete Fourier Transform

When the summation in (3.79) is expanded for N ¼ 4; we get ~x3 ½n ¼ ~x1 ½0~x2 ½n þ ~x1 ½1~x2 ½n  1 þ ~x1 ½2~x2 ½n  2 þ ~x1 ½3~x2 ½n  3:

ð3:80Þ

One period of ~x2 ½n; ~x2 ½n  1; ~x2 ½n  2; and ~x2 ½n  3 for 0  n  3 can be calculated using rotate right operation yielding ~x2op ½n ¼ ½ 1 1 1 1  ~x2op ½n  1 ¼ ½ 1 1 1 1  ~x2op ½n  2 ¼ ½ 1 1 1 1  ~x2op ½n  3 ¼ ½ 1 1 1 1 :

ð3:81Þ

Substituting (3.81) into (3.80), one period of ~x3 ½n is calculated as ~x3op ½n ¼ ~x1 ½0~x2op ½n þ ~x1 ½1~x2op ½n  1 þ ~x1 ½2~x2op ½n  2 þ ~x1 ½3~x2op ½n  3 yielding ~x3op ½n ¼ 1  ½ 1

1

1

þ2  ½1

1

1

1  þ 1  ½ 1

1 1

1  1  ½1

1

1

1

1 

which can be simplified as ~x3op ½n ¼ ½ 1

3

1 3 :

ð3:82Þ

Using (3.82), the periodic convolution result can be written as ~x3 ½n ¼ ½    1

3 13

1 |{z}

3

13

1 3

1

3

  :

n¼0

3.3.1

Alternative Method to Compute the Periodic Convolution

The periodic convolution expression ~x3 ½n ¼

N 1 X m¼0

~x1 ½m~x2 ½n  m

ð3:83Þ

3.3 Convolution of Periodic Digital Signals

167

can be computed for n ¼ 0; 1; . . .; N  1 as n ¼ 0; n ¼ 1; n ¼ 2;

~x3 ½0 ¼ ~x3 ½1 ¼ ~x3 ½2 ¼

N1 P

~x1 ½m~x2 ½m

m¼0 N1 P

~x1 ½m~x2 ½1  m

m¼0 N1 P

~x1 ½m~x2 ½2  m

m¼0

.. .

n ¼ N  1;

~x3 ½N  1 ¼

N1 P

~x1 ½m~x2 ½ðN  1Þ  m:

m¼0

Now let’s consider ~x3 ½0 ¼

N 1 X

~x1 ½m~x2 ½m

ð3:84Þ

m¼0

when expanded for N ¼ 3; we get ~x3 ½0 ¼ ~x1 ½0~x2 ½0 þ ~x1 ½1~x2 ½1 þ ~x1 ½2~x2 ½2 þ ~x1 ½3~x2 ½3

ð3:85Þ

Since ~x3 ½n ¼ ~x3 ½n þ 4, we have ~x2 ½1 ¼ ~x2 ½3;

~x2 ½2 ¼ ~x2 ½2;

~x2 ½3 ¼ ~x2 ½1:

ð3:86Þ

Using (3.86) in (3.85), we obtain ~x3 ½0 ¼ ~x1 ½0~x2 ½0 þ ~x1 ½1~x2 ½3 þ ~x1 ½2~x2 ½2 þ ~x1 ½3~x2 ½1

ð3:87Þ

which can be written as the dot product of the vectors ½ ~x1 ½0 ~x1 ½1 ~x1 ½2 ~x1 ½3 and ½ ~x2 ½0 ~x2 ½0 ~x2 ½2 ~x2 ½1  where it is clear that the vector ½ ~x2 ½0 ~x2 ½3 ~x2 ½3 ~x2 ½1  can be obtained from one period of ~x2 ½n via rotate inside operation. Hence we can write N 1 X

~x1 ½m~x2 ½m ! ~x3 ½0 ¼ ~x1op ½m ~x2op ½m

ð3:88Þ

~x1 ½m~x2 ½1  m ! ~x3 ½1 ¼ ~x1op ½m ~x2op ½1  m:

ð3:89Þ

~x3 ½0 ¼

m¼0

~x3 ½1 ¼

N 1 X m¼0

168

3

Discrete Fourier Transform

Equation (3.89) can be written as ~x3 ½1 ¼

N 1 X



~x1 ½m~x2 ½1  m ! ~x3 ½1 ¼ ~x1op ½m RR ~x2op ½m

ð3:90Þ

m¼0

and in a similar manner ~x3 ½2 ¼

N 1 X



~x1 ½m~x2 ½2  m ! ~x3 ½2 ¼ ~x1op ½m RR ~x2op ½1  m

ð3:91Þ



~x1 ½m~x2 ½3  m ! ~x3 ½3 ¼ ~x1op ½m RR ~x2op ½2  m

ð3:92Þ

m¼0

~x3 ½3 ¼

N 1 X m¼0

.. . ~x3 ½N  1 ¼

N 1 X

~x1 ½m~x2 ½N  1  m !

m¼0



~x3 ½N  1 ¼ ~x1op ½m RR ~x2op ½N  2  m

ð3:93Þ

Example 3.15 The periodic signals ~x1 ½n and ~x2 ½n with period ¼ 4 are shown in Fig. 3.17. Calculate their 4-point periodic convolution using alternative periodic convolution method.

~

x1[n]

2

1

1

4

3

2

1

2

1

1

0

1

1

2

3

4

n

1

1 ~

x2 [n]

4 1

1

1

3

2

1 1

0

1

1

1

2

3

1

Fig. 3.17 The periodic signals ~x1 ½n and ~x2 ½n for Example 3.15

1

4

1

n

3.3 Convolution of Periodic Digital Signals

169

Solution 3.15 When the periodic convolution expression ~x3 ½n ¼

N 1 X

~x1 ½m~x2 ½n  m

m¼0

is calculated for n ¼ 0; 1; . . .; N  1, we get ~x3 ½0 ¼ ~x1op ½m ~x2op ½m ~x3 ½1 ¼ ~x1op ½m ~x2op ½1  m

~x3 ½2 ¼ ~x1op ½m RR ~x2op ½1  m

~x3 ½3 ¼ ~x1op ½m RR ~x2op ½2  m :

ð3:94Þ

One period of ~x2 ½n for 0  n  3 is ~x2op ½m ¼ ½ 1

1

1

1 :

Then ~x2op ½m ¼ ½ 1 1 1 1  ~x2op ½1  m ¼ RRð~x2op ½mÞ ! RRð~x2op ½mÞ ¼ ½ 1 1 1 1  ~x2op ½2  m ¼ RRð~x2op ½1  mÞ ! RRð~x2op ½mÞ ¼ ½ 1 1 1 1  ~x2op ½3  m ¼ RRð~x2op ½2  mÞ ! RRð~x2op ½mÞ ¼ ½ 1 1 1 1 

ð3:95Þ

and ~x1op ½m ¼ ½ 1 1

1

2 :

ð3:96Þ

Using (3.95) and (3.96) in (3.94), we can calculate the periodic convolution values as ~x3 ½0 ¼ ½ 1 1 1 2  ½ 1 1 1 1  ! ~x3 ½0 ¼ 1ð1Þ þ 1ð1Þ þ ð1Þ1 þ 2  1 ! ~x3 ½3 ¼ 1 ~x3 ½1 ¼ ½ 1 ~x3 ½2 ¼ ½ 1 ~x3 ½3 ¼ ½ 1

1 1 1

1  ! ~x3 ½1 ¼ 3 1 2  ½ 1 1 1 1  ! ~x3 ½2 ¼ 1 1 2  ½ 1 1 1 1  ! ~x3 ½3 ¼ 3 1 2  ½ 1

1

1

170

3

Discrete Fourier Transform

Hence, ~x3op ½n ¼ ½ 1

3

1 3 :

Then the periodic convolution result becomes as ~x3 ½n ¼ ½  

1

3

1

3

1 |{z}

3 1

3

1 3

3

1

  :

n¼0

3.4

Sampling of Fourier Transform

The Fourier transform Xn ðwÞ of a non-periodic digital signal x½n is a continuous function of w and it is periodic with period 2p, i.e., Xn ðwÞ ¼ Xn ðw þ 2pÞ . Example 3.16 The Fourier transform of the signal x½n ¼ 12 d½n þ 1 þ 12 d½n  1 is calculated as Xn ðwÞ ¼

1 X

x½nejwn

n¼1 1  X

 1 1 d½n þ 1 þ d½n  1 ejwn ¼ 2 2 n¼1

1 ¼ ejw þ ejw 2 ¼ cosðwÞ:

ð3:97Þ

The aperiodic digital signal x½n and its Fourier transform is shown in Fig. 3.18. Let’s generate the periodic signal ~x½n with period N from x½n via ~x½n ¼

1 X

x½n  lN:

ð3:98Þ

l¼1

Xn(w)

x[n] 1 2 1

1

x[n] e jwn

X n (w) n

1

2

0

3 2

2

1

3 2

2 1

Fig. 3.18 The aperiodic digital signal x½n in Example 3.16 and its Fourier transform

2

w

3.4 Sampling of Fourier Transform

171

The Fourier series coefficients of the periodic signal ~x½n in (3.98) are obtained from the Fourier transform of x½n, i.e., Xn ðwÞ, via sampling operation in frequency domain as in ~ ½k ¼ Xn ðwÞj X w¼kws

ð3:99Þ

where ws ¼ 2p N is the sampling period in radian unit. Example 3.17 ~x½n is a periodic signal with period N ¼ 4, and we have x½n ¼  1 for one period of this signal. In addition, the periodic signal can be obtained from its one period via 1 1 2 d½ n þ 1 þ 2 d½ n

~x½n ¼

1 X

x½n  lN:

ð3:100Þ

l¼1

Find the Fourier series coefficients of ~x½n using Xn ðwÞ the Fourier transform of x½n: Solution 3.17 In Example 3.17, we found the Fourier transform of x½n ¼  1 as

1 1 2 d½ n þ 1 þ 2 d½ n

Xn ðwÞ ¼ cosðwÞ: ~ ½k, of ~x½n can be obtained via sampling The Fourier series coefficients, i.e., X operation in frequency using ~ ½k ¼ Xn ðwÞj X w¼kws

ð3:101Þ

2p p where ws ¼ 2p N ! ws ¼ 4 ! ws ¼ 2. Hence (3.101) yields

  ~ ½k ¼ cosðwÞjw¼kp ! X ~ ½k  ¼ cos kp : ~ ½k ¼ Xn ðwÞjw¼kw ! X X s 2 2

ð3:102Þ

The graphical illustration of the sampling operation in frequency domain is explained in Fig. 3.19.

172

3

X n (w)

x[n]

1 2 1

Discrete Fourier Transform

1

x[n] e jwn

X n (w) n

1

n

0

3 2

2

3 2

2

2

~

ws

x[n 4l ] l

Sample the signal in frequency domain at every multiple of ws

2

~

x[n]

X n (w)

~

X [k ]

1 2

4

1

w

1

1

x[n]

2

1

1

4

n

0

3 2

2

3 2

2

2

X [k ] [ 1

0

w

1

1 ~

2

1

0

1

0

1

0

]

k 0

Fig. 3.19 Fourier series coefficients are obtained from Fourier transform via sampling operation

3.5

Discrete Fourier Transform

~ ½k, is a periodic function which can have The Fourier series coefficients, i.e., X ~ ½k satisfy X ~ ½k  ¼ X ~ ½k þ N  complex or real values. The Fourier series coefficients X where N is the period of the digital signal ~x½n. The periodic signal ~x½n with period N has the Fourier series coefficients ~ ½k  ¼ X

N 1 X

~x½nej N kn 2p

ð3:103Þ

n¼0

and for 0  n\N, ~x½n ¼ x½n where x½n is one period of ~x½n. Then (3.103) can be written as ~ ½k  ¼ X

N 1 X

x½nej N kn 2p

ð3:104Þ

n¼0

which is also a periodic signal with the same period as the time domain signal ~x½n. ~ ½k  Let’s consider one period of X

3.5 Discrete Fourier Transform

173

 X ½k  ¼

~ ½k  if 0  k\N X 0 otherwise

ð3:105Þ

which is called the discrete Fourier transform of x½n. Thus, N  point discrete Fourier transform of x½n is defined as X ½k  ¼

N 1 X

x½nej N kn ; 2p

0  k\N:

ð3:106Þ

n¼0

Similarly, N  point inverse Fourier transform is defined as x ½ n ¼

N1 1X 2p X ½kej N kn ; N k¼0

0  n\N:

A more general definition for N-point DFT is X ½k  ¼

X

x½nej N kn ; 2p

k; N:

ð3:107Þ

n;N

and for the N-point inverse DFT, a more general definition is x½n ¼

1X 2p X ½k ej N kn ; N k;N

n; N:

ð3:108Þ

In addition, Fourier series coefficients of a periodic signal can be obtained from the Fourier transform of its one period using ~ ½k ¼ Xn ðwÞj X w¼kws

ws ¼

2p : N

ð3:109Þ

And using the definition  X ½k  ¼

~ ½k  if 0  k\N X 0 otherwise

ð3:110Þ

we can write X ½k  ¼ Xn ðwÞjw¼kws ;

ws ¼

2p ; 0  k\N N

ð3:111Þ

which means that the discrete Fourier transform of x½n is nothing but a mathematical sequence obtained from one period of Xn ðwÞ via sampling operation in frequency domain, and the sampling period is chosen as ws ¼ 2p N.

174

3

Discrete Fourier Transform

Example 3.18 Find the discrete Fourier transform of x½n ¼ ½ 1 1

1

2 :

Solution 3.18 For the given signal if the DFT formula X ½k  ¼

41 X

x½nej 4 kn ; 2p

0  k\4

ð3:112Þ

n¼0

is expanded, the coefficients are found as X ½k  ¼ x½0 e0 þ x½1 ej 4 k þ x½2 ej 4 k2 þ x½3 ej 4 k3 : |{z} |{z} |{z} |{z} 2p

1

2p

2p

1

1

ð3:113Þ

2

When (3.113) is simplified, we obtain X ½k ¼ 1 þ ej2k  1ejpk þ 2ej 2 k p

3p

ð3:114Þ

Evaluating (3.114), i.e., X ½k, for k ¼ 0; 1; 2; 3, we get X ½ 0 ¼ 3

X ½1 ¼ 2 þ j X ½2 ¼ 3

X ½3 ¼ 2  j

which can be written in short as X ½k  ¼ ½ 3

2 þ j 3

2  j :

Example 3.19 Find the aperiodic digital signal whose DFT coefficients are given as X ½k  ¼ ½ 3

2 þ j 3

2  j :

Solution 3.19 Using X ½k  in inverse DFT formula x½n ¼

41 1X 2p X½kej 4 kn ; 4 k¼0

0  n\4

ð3:115Þ

we obtain 0

1

1B 2p 2p 2p C x½n ¼ @ X ½0 ej0 þ X ½1 ej 4 1n þ X ½2 ej 4 2n þ X ½3 ej 4 3n A: |{z} |{z} |{z} 4 |{z} 3

2þj

3

2j

ð3:115Þ

3.5 Discrete Fourier Transform

175

When (3.115) is simplified, we get x ½ n ¼

 1 p 3p 3 þ ð2 þ jÞej2n  3ejpn þ ð2  jÞej 2 n 4

ð3:116Þ

Evaluating (3.116), i.e., x½n, for n ¼ 0; 1; 2; 3, we obtain, x ½ 0 ¼ 1 x ½ 1 ¼ 1

x½2 ¼ 1

x ½ 3 ¼ 2

which can be written in short as x½n ¼ ½ 1 1

1

2 :

Note: Remember that ejh ¼ cosðhÞ þ j sinðhÞ. Example 3.20 Find the discrete Fourier transform of the signal shown in Fig. 3.20. Solution 3.20 Using the DFT formula X ½k  ¼

X

x½nej N kn 2p

k; N

n;N

for N ¼ 3, we obtain X ½k  ¼

1 X

x½nej N kn 2p

 1  k  1:

ð3:117Þ

n¼1

When (3.117) is expanded, we get X ½k ¼ x½1 ej 3 kð1Þ þ x½1 ej 3 k1 ; |fflffl{zfflffl} |{z} 2p

2p

1=2

1  k  1

ð3:118Þ

1=2

which is simplified as 

 2p k ; X ½k ¼ cos 3

1  k  1:

ð3:119Þ

x[n ]

Fig. 3.20 Aperiodic signal for Example 3.20

1 2 1

1

n

176

3

Discrete Fourier Transform

From (3.115) DFT coefficients can be calculated as k ¼ 1 ! X ½1 ¼ 

1 2

k ¼ 0 ! X ½0 ¼ 1 k ¼ 1 ! X ½1 ¼ 

1 2

That is, X ½k  ¼ ½

 12

1 1 |{z}

 12

:

k¼0

Note: For the previous example, discrete Fourier transform is calculated for N ¼ 3 which is equal to the length of the aperiodic sequence x½n. Hence, if it is not clearly mentioned, the default length of the DFT computation is the same as the length of the aperiodic sequence x½n. Example 3.21 DFT coefficients of an aperiodic signal are given as X ½k  ¼ ½

 12

1 1 |{z}

 12

:

ð3:120Þ

k¼0

Find x½n whose DFT coefficients are X½k. Solution 3.21 If we use inverse DFT formula x½n ¼

1X 2p X ½k ej N kn ; N k;N

n; N

for the given signal, we get x½n ¼

1 1X 2p X ½kej 3 kn ; 3 k¼1

1  n  1:

ð3:121Þ

When the summation term in (3.121) is expanded, we obtain 0

1

1B 2p 2p C x½n ¼ @X ½1 ej 3 n þ X ½0 e0 þ X ½1 ej 3 n A |{z} |{z} 3 |fflffl{zfflffl} 1=2

1

1=2

3.5 Discrete Fourier Transform

177

which is simplified as x ½ n ¼

  1 1 2p 1 2p  ej 3 n þ 1  ej 3 n : 3 2 2

ð3:122Þ

Let’s evaluate (3.122), i.e., x½n, for n ¼ 1; 0; 1. We first calculate for n ¼ 1 as x½1 ¼

  1 1 2p 1 2p  ej 3 ð1Þ þ 1  ej 3 ð1Þ 3 2 2

which is simplified as     1 2p 1 1 1 x½1 ¼ cosð Þ þ 1 ! x½1 ¼ þ 1 ! x½1 ¼ 3 3 3 2 2 and for n ¼ 0, we have   1 1 0 1 0 x ½ n ¼  e þ 1  e ! x ½ n ¼ 0 3 2 2 and finally for n ¼ 1, we get x ½ 1 ¼

  1 1 2p 1 2p  ej 3 ð1Þ þ 1  ej 3 ð1Þ 3 2 2

which is simplified as x ½ 1 ¼

    1 2p 1 1 1 cosð Þ þ 1 ! x½1 ¼ þ 1 ! x ½ 1 ¼ : 3 3 3 2 2

Thus the signal x½n has the values x½1 ¼ 12 x½0 ¼ 0 x½1 ¼ 12 which is written in more compact form as 1

x½n ¼ ½ 2

0 |{z}

1 2 :

n¼0

Question: For the previous example if we evaluate   1 1 j2pn 1 j2pn 3 3  e þ1  e x ½ n ¼ 3 2 2

ð3:123Þ

178

3

Discrete Fourier Transform x[n ]

Fig. 3.21 Aperiodic signal for Example 3.22

1 2 1

1

n

for n ¼ 0; 1; and 2, we obtain 0 x½n ¼ ½ |{z}

1 2

1 2 :

ð3:124Þ

n¼0

When (3.123) and (3.124) are compared to each other, we see that (3.124) can be obtained from (3.123) by rotate left or rotate right operations. Example 3.22 Find the 8-point discrete Fourier transform of the signal in Fig. 3.21. Solution 3.22 Although the length of the aperiodic signal equals to 2, the DFT will be calculated for 8-points. For this reason, we first pad the signal by zeros so that its length equals to 8. So the finite length signal becomes as x½n ¼ ½1

0 |{z}

1 0

0

0

0 0:

n¼0

And the 8-point DFT is computed using X ½k  ¼

6 X

x½nej 8 kn ; 2p

1  k  6:

ð3:125Þ

n¼1

When the summation in (3.125) is expanded, we get X ½k  ¼

    1 1 2p 2p 2p 2p  ej 8 kð1Þ þ 0  ej 8 k0 þ  ej 8 k1 þ 0  ej 8 k2 2 2 þ 0  ej 8 k3 þ 0  ej 8 k4 þ 0  ej 8 k5 þ 0  ej 8 k6 2p

2p

2p

2p

which is simplified as X ½k  ¼

 1  j2pk 2p e 8 þ ej 8 kn : 2

ð3:126Þ

3.5 Discrete Fourier Transform

179

Equation (3.126) can be written in terms of cosðÞ function as  X ½k ¼ cos

 2p k ; 8

1  k  6:

ð3:127Þ

And when the Fourier series coefficients in (3.127) are explicitly calculated, we obtain

 X ½k ¼ cos  2p 8

cosð0Þ

cos

2p

8

cos

4p

8

cos

6p

8

cos

8p

8

cos

10p

8

cos

12p 8

which is simplified as X ½k  ¼ ½0:7071

1 |{z}

0:7071

0

0:7071

1

0:7071

0:

k¼0

Example 3.23 DFT coefficients are complex numbers. And those complex coefficients have magnitude and phase values. For the DFT coefficients X ½k  ¼ ½ 3

2 þ j 3 þ j

2  j

find jX ½kj, i.e., magnitudes of the DFT coefficients, and \X ½k , i.e., phase information of DFT coefficients. Solution 3.23 For the complex number x ¼ a þ bj the magnitude and phase information is calculated as pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi j xj ¼ ða2 þ b2 Þ;

  b \ tan : a 1

ð3:128Þ

Using (3.128) the magnitude and phase of each DFT coefficient is calculated as pffiffiffiffiffiffiffiffiffiffiffiffiffiffi j X ½ 0 j ¼ 32 þ 02 ! 3 pffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffi j X ½ 1 j ¼ 22 þ 12 ! 5 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffi jX ½2j ¼ ð3Þ2 þ 12 ! 10 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffi jX ½3j ¼ 22 þ ð1Þ2 ! 5

\X ½0 ¼ tan1 03 ! 0 \X ½1 ¼ tan1 12 ! 0:15p \X ½2 ¼ tan1  13 ! 0:1p \X ½3 ¼ tan1  12 ! 0:15p

Magnitude and phase values are plotted in Fig. 3.22. Example 3.24 One period of the discrete time Fourier transform of the non-periodic signal x½n is given in Fig. 3.23. Using the given Fourier transform graph: (a) Find the 4-point DFT of x½n. (b) Find the 8-point DFT of x½n. (c) Find the 16-point DFT of x½n.

180

3

Discrete Fourier Transform

| X [k ] | 10

3

5

0

5

2

1

k

3

X [k ]

0.15

1

0

k

2

3

0 .1

0.15

Fig. 3.22 Magnitude and phase plot of DFT coefficients in Example 3.23

X n (w) 3

2 1.25 0.75 0.32 0

π 4

2π 4

3π 4

4π 4

5π 4

6π 4

7π 4

8π 4

w

Fig. 3.23 One period of the discrete time Fourier transform of a non-periodic signal

Solution 3.24 (a) DFT coefficients are obtained by sampling of Xn ðwÞ in frequency domain. That is, 2p : ð3:129Þ X ½k ¼ Xn ðwÞjw¼kws ws ¼ N Since N ¼ 4, we take 4 samples from one period of Xn ðwÞ. The sampling period is ws ¼

2p 2p ! ws ¼ : 8 4

The sampling operation is illustrated in Fig. 3.24.

3.5 Discrete Fourier Transform

181

X n (w) w kws

3

2 4

ws

2

1.25 0.75 0.32 0

k

2 4 k 1

0

8 4

6 4 k 3

4 4 k 2

w

Fig. 3.24 Sampling of the Fourier transform for N ¼ 4

Considering Fig. 3.24, the DFT coefficients can be written as X ½k  ¼ ½ 0

0:75 :

3 1:25

(b) For N ¼ 8, we take 8 samples from one period of Xn ðwÞ. The sampling period is ws ¼

2p p ! ws ¼ : 8 4

The sampling operation for N ¼ 8 is illustrated in Fig. 3.25. Thus the DFT coefficients obtained in Fig. 3.25 can be written as a mathematical sequence as X ½k  ¼ ½ 0 2

3

3

1:25

0:75

0:75

0:32 :

X n (w) 3

w kws

ws

2 8

2

1.25 0.75 0.32 0

2 4

4 k

0

k

1

k

3 4 2

k

4 4 3

k

5 4 4

Fig. 3.25 Sampling of the Fourier transform for N ¼ 8

k

5

k

8 4

7 4

6 4 6

k

7

w

182

3

Discrete Fourier Transform

Exercise: The aperiodic signal is given as x½n ¼ d½n þ d½n  1. (a) Find the Fourier transform of x½n, i.e., Xn ðwÞ ¼ ? (b) Find jXn ðwÞj P and \Xn ðwÞ. (c) If ~x½n ¼ 1 x½n and using Xn ðwÞ, find the Fourier series l¼1 x½n  4l, draw ~ ~ ½k  ¼ ? coefficients of ~x½n, i.e., X (d) Find 4-point DFT of x½n

3.5.1

Aliasing in Time Domain

When we study sampling theorem, we have seen that during sampling operation if we do not take sufficient number of samples from analog signal, we cannot perfectly reconstruct analog signal at the receiver side from its digital samples. And the effect of this situation is seen as aliasing or overlapping in frequency domain. We have seen that DFT coefficients of a non-periodic digital signal x½n are nothing but the samples taken from one period of its Fourier transform, for instance, samples taken for 0  w\2p. We can reconstruct the digital signal x½n from its DFT coefficients using x r ½ n ¼

N 1 1X 2p X ½kej N kn ; N k¼0

0  n\N:

ð3:130Þ

Now we ask the question: Is xr ½n always equal to x½n ? If not always, then what is the criteria for xr ½n to be equal to x½n ? We know that N-point DFT coeffcients of x½n equals to the one period of the DFS coefficients of the periodic signal ~x½n, and the relation between x½n and ~x½n can be stated as ~x½n ¼

1 X

x½n  kN :

ð3:131Þ

k¼1

Let the length of the digital signal x½n be M. If M [ N, then the shifted successor signals x½n  kN  overlap each other. And when the shifted signals are summed, one period of ~x½n is not equal to x½n anymore. This means that using the inverse DFT operation, x½n cannot be obtained exactly. The amount of distortion in the reconstructed signal depends on the overlapping amount. Example 3.25 For x½n ¼ ½ 1 1 ~x½n ¼

1  and N ¼ 2, calculate 1 X

x½n  kN :

k¼1

Find one period of ~x½n and compare it to x½n.

3.5 Discrete Fourier Transform

183 x [n 4]

x [n 4]

1

6

1

4

5

3

1 1

x [n 2]

2

1

1

1

1

1 1

3

4

5

6

x [n 2]

1 1 0

1

1 1 2

x[n ]

1 1

n

Fig. 3.26 Shifted signals ~

x[n] 0 6

1 5

0

1

0

1

0

1

0

1

0

1

4

3

2

1

0

1

2

3

4

5

6

n

Fig. 3.27 Sum of the shifted signals in Fig. 3.26

Solution 3.25 The shifted signals are shown in Fig. 3.26. The sum of the shifted signals in Fig. 3.26 yields the signal in Fig. 3.27. As it is seen from Fig. 3.27, one period of ~x½n is [0 1] which is totally different than x½n ¼ ½ 1 1 1 . Example 3.26 x½n ¼ ½ 1 1 1 , calculate 2-point DFT of x½n and using 2-point DFT coefficients, calculate x½n using the inverse DFT formula and comment on the results. Solution 3.26 2-point DFT coefficients of x½n ¼ ½ 1 using X ½k  ¼

1 X

x½nej 2 kn ; 2p

1

1  can be calculated

0k1

n¼0

yielding X2 ½k ¼ ½ 0

2 :

and proceeding in a similar manner 3-point DFT coefficients can be found as X3 ½k ¼ ½ 1

2

2 :

If we use the 2-point inverse DFT formula for X2 ½k  x ½ n ¼

1 1X 2p X2 ½k ej 2 kn ; 2 k¼0

0n1

184

3

Discrete Fourier Transform

the aperiodic signal is found as x½n ¼ ½ 1

1

which is truncated version of x½n ¼ ½ 1

3.5.2

1 :

1

Matrix Representation of DFT and Inverse DFT

Before generalizing the concept, let’s consider 3-point DFT of an aperiodic sequence X ½k  ¼

2 X

x½nej 3 kn ; 2p

0k2

ð3:132Þ

n¼0

When the summation in (3.132) is expanded for each k value, we obtain the following equations X ½0 ¼ x½0e0 þ x½1e0 þ x½2e0 X ½1 ¼ x½0e0 þ x½1ej 3 þ x½2ej 3 X ½2 ¼ x½0e þ x½1e 0

2p

4p

j4p 3

j8p 3

þ x½2e

ð3:133Þ :

The equation set in (3.133) can be written as 3 X½0 4 X½1 5 ¼ ½ x½0 X½2

2

2

x½1

e0 4 x½2   e0 e0

e0

e e

j2p 3 j4p 3

e0

e e

j4p 3 j8p 3

3 5

which can be expressed in short as  ½ k ¼  X x½n  EN ;

N ¼ 3:

ð3:134Þ

From (3.134)  x½n can be written as  ½k  E1 :  x½n ¼ X N

ð3:135Þ

3.5 Discrete Fourier Transform

185

In a similar manner, the inverse 3-point DFT formula can be written in matrix form. Expanding x½n ¼

2 1X 2p X ½k ej 3 kn ; 3 k¼0

0n2

ð3:136Þ

we get

1 X ½0e0 þ X ½1e0 þ X ½2e0 3  1 2p 4p X ½0e0 þ X ½1ej 3 þ X ½2ej 3 x½1 ¼ 3  1 4p 8p X ½0e0 þ X ½1ej 3 þ X ½2ej 3 : x½2 ¼ 3

x½0 ¼

ð3:137Þ

The equation set in (3.137) can be written in matrix form as 2

3 x½0 4 x½1 5 ¼ 1  ½ X ½0 3 x½2

2

e0 4 X ½ 1 X ½ 2   e 0 e0

e0 2p ej 3 4p ej 3

3 e0 4p ej 3 5 : 8p ej 3

ð3:138Þ

When (3.138) is compared to (3.139)  ½k  E1  x½n ¼ X N

ð3:139Þ

we obtain E1 N ¼

1 E : N N

ð3:140Þ

Note: E N is the conjugate of EN . If e ¼ a þ jb then conjugate of e is e ¼ a  jb and if e ¼ ejh then e ¼ ejh .

3.5.3

Properties of the Discrete Fourier Transform

Since there is a close relationship between discrete Fourier series coefficients of a periodic signal and the discrete Fourier transform of its one period, it is logical to review the properties of the discrete Fourier series coefficients of a periodic signal. For the three periodic signals ~x½n ! Periodic with period N ~x1 ½n ! Periodic with period N ~x2 ½n ! Periodic with period N

186

3

Discrete Fourier Transform

let’s denote the Fourier series coefficients by ~ ½k ! Periodic with period N X ~1 ½k ! Periodic with period N X ~2 ½k  ! Periodic with period N: X And the correspondence between signals and their DFS coefficients are shown as DFS ~ ~x½n $ X½k DFS ~1 ½k  ~x1 ½n $ X DFS ~2 ½k : ~x2 ½n $ X

Properties Linearity: DFS ~1 ½k  þ bX ~2 ½k  a~x1 ½n þ b~x2 ½n $ aX

Duality: ~ ½n DFS X $ N~x½k Shifting in time: ~ ~x½n  m $ ej N km X½k DFS

2p

Shifting in frequency: 2p DFS ~  l ej N ln~x½n $ X½k

Convolution in time domain: N 1 X

DFS ~ 1 ½k X ~2 ½k  ~x1 ½m~x2 ½n  m $ X

m¼0

Convolution in frequency domain: DFS

~x1 ½n~x2 ½n $

N 1 1X ~1 ½mX ~2 ½k  m X N k¼0

3.5 Discrete Fourier Transform

187

Conjugate: DFS ~ ½k ~x ½n $ X

Real part DFS: DFS 1

Ref~x½ng $

2

~ ½k  þ X ~ ½k  X

Imaginary part DFS: DFS 1

jImf~x½ng $

2

~ ½k   X ~ ½k X



Real part: 1 DFS ~ ð~x½n þ ~x ½nÞ $ RefX½kg 2 Imaginary part: 1 DFS ~ ð~x½n  ~x ½nÞ $ jImfX½kg 2 For real ~ x½n, we have the following properties Conjugate: ~ ½k  ¼ X ~ ½k X Real DFT coefficients:

 ~ ½k  ¼ RefX ~ ½kg Re X Imaginary DFT coefficients:

 ~ ½k  ¼ ImfX ~ ½kg Im X Absolute value:     X ~ ½k  ¼ X ~ ½k  Phase value: ~ ½k ¼ \X ~ ½k \X

188

3

Discrete Fourier Transform

Real part: 1 DFS ~ ð~x½n þ ~x½nÞ $ RefX½kg 2 Imaginary part: 1 DFS ~ ð~x½n  ~x½nÞ $ jImfX½kg 2 Note: If x½n ¼ a½n þ jb½n, then x ½n ¼ a½n  jb½n

3.5.4

Circular Convolution

The discrete Fourier transform of an aperiodic sequence x½n with length N equals to the one period of the Fourier series coefficients of the periodic signal ~x½n obtained from x½n as ~x½n ¼

1 X

x½n  kN

k¼1

and the relation between DFT coefficients of x½n and one period of Fourier series coefficients of the periodic signal ~x½n is given as  X ½k  ¼

~ ½k  if 0  k  N  1 X 0 otherwise:

Let’s denote one period of ~x½n for 0  n  N  1 by x½ðnÞN . It is clear that if the  length of x½n is N then x ðnÞN ¼ x½n. However, if the length of x½n is a number other than N then  x ðnÞN 6¼ x½n: If not indicated otherwise, we will assume that the length of x½n and period of ~x½n are equal to each other. Properties x1 ½n ! Aperiodic signal with length N1 x2 ½n ! Aperiodic signal with length N2 N ¼ maxfN1 ; N2 g

3.5 Discrete Fourier Transform

189

x 1 ½ n

NpointDFT

$

X1 ½k

x 2 ½ n

NpointDFT

X2 ½k

$

Linearity: DFT

ax1 ½n þ bx2 ½n $ aX1 ½k þ aX2 ½k  Circular Shifting:  DFT 2p x ðn  mÞN $ ej N km X ½k  Duality: DFT

x½n $ X ½k   DFT X ½n $ Nx ðkÞN Symmetry:  DFT x ½n $ X ðk ÞN  DFT X ðnÞN $ X ½k Symmetry property leads to the following properties DFT

Refx½ng $ Xep ½k; DFT

jImfx½ng $ Xop ½k ;

ep :even part op :odd part

DFT

xep ½n $ RefX ½k g DFT

xop ½n $ jImfX ½kg Circular Convolution: DFT

x1 ½n $ X1 ½k  DFT

x2 ½n $ X2 ½k 

190

3

Discrete Fourier Transform

If Y ½k  ¼ X1 ½k X2 ½k then y ½ n ¼

N 1 X

x1 ½mx2 ½ðn  mÞN 

m¼0

or y ½ n ¼

N 1 X

x2 ½mx1 ½ðn  mÞN :

m¼0

And the expression N 1 X

x1 ½mx2 ½ðn  mÞN 

m¼0

is called the circular convolution of x1 ½n and x2 ½n and denoted by  Example 3.27 What does x ðnÞ5 0  n  4 mean? Solution 3.27 x½ðnÞ5  equals to one period of ~x½n in the interval 0  n  4, i.e.,  x ðnÞ5 ¼ ~x½n 0  n  4 and ~x½n ¼

1 X

x½n  5l:

l¼1

Note: We assumed that the length of x½n and period of ~x½n are equal to each other.  Example 3.28 If x½n ¼ ½ 1 1 1 0:5 1 , find x ðnÞ5 0  n  4. Solution 3.28 x½ðnÞ5  equals to ~y½n ¼ ~x½n for 0  n  4 and ~x½n is given as

3.5 Discrete Fourier Transform

191 1 X

~x½n ¼

x½n  5l:

l¼1

One period of ~x½n in the interval 0  n  4 is found by employing rotate inside operation on one period of ~x½n, i.e., on x½n. That is  x ðnÞ5 ¼ RI ðx½nÞ which can be calculated as  x ðnÞ5 ¼ ½ 1

1

0:5 1

1 :

 Example 3.29 If x½n ¼ ½ 1 1 1 0:5 1 , find x ð1  nÞ5 .  Solution 3.29 x ð1  nÞ5 equals to ~x½1  n for 0  n  4 and ~x½n is calculated as 1 X

~x½n ¼

x½n  5l:

l¼1

One period of ~x½1  n is obtained by rotating one period of ~x½n to the right by ‘1’ unit. That is  

x ð1  nÞ5 ¼ RR x ðnÞ5 : Using the result of the previous example, i.e.,  x ðnÞ5 ¼ ½ 1

1

0:5

1

1

 we can calculate x ð1  nÞ5 via  

x ð1  nÞ5 ¼ RR x ðnÞ5 which yields  x ð 1  nÞ 5 ¼ ½ 1

1

1

0:5

1 :

  Note: x ð2  nÞ5 is obtained by rotating x ð1  nÞ5 to the right by ‘1’ unit.   And x ð1  nÞ5 is obtained by rotating x ðnÞ5 to the left by ‘1’ unit.  Example 3.30 If x½n ¼ ½ 1 1 1 1 , find x ðnÞ3 .  Solution 3.30 x ðnÞ3 ¼ ~x½n for 0  n  3 and ~x½n is obtained as

192

3

1

6

x [n 6] 1 1

4

5

1 1

x [n 3] 1 1

3

2

1 1 0

1

x [n] 1 1 1 1 1 2 3 1 1 0

Discrete Fourier Transform

x [n 3] 1 1

1 1

4

6

5

x [n 6] 1 1

1

n

~

x [ n] 0

1

1

0

1

1

0

1

1

0

1

1

6

5

4

3

2

1

0

1

2

3

4

5

6

n

Fig. 3.28 Shifted replicas of x½n and calculation of ~x½n

~x½n ¼

1 X

x½n  3l:

l¼1

P Since the length of x½n is 4, the shifted successor copies in 1 l¼1 x½n  3l overlap with each other. For this reason, one period of ~x½n is not equal to x½n anymore. It should be calculated explicitly. This calculation is explained in Fig. 3.28.  x ðnÞ3 for 0  n  3 equals to one period of ~x½n and from Fig. 3.28, it is found as

~xop ½n ¼ ½ 0

1

1

1

1 :

 which is denoted by x ðnÞ3 , that is,  x ð nÞ 3 ¼ ½ 0

 And x ðnÞ3 which is equal to one period of ~x½n can be found using the rotate inside operation as 

x ðnÞ3 ¼ RI ~xop ½n yielding  x ðnÞ3 ¼ ½ 0

1

1 :

 Exercise: For the previous example find x ð2  nÞ3 . Example 3.31 If x½n ¼ ½ 0:5

0:5

0:5

1

1 , find x½ðn  2Þ5 .

Solution 3.31 x½ðn  2Þ5  equals to ~x½n  2 for 0  n  4 and ~x½n is obtained as

3.5 Discrete Fourier Transform

193 1 X

~x½n ¼

x½n  lN

l¼1

where N ¼ 5. Since the length of x½n equals to the period value of the ~x½n, then ~x½n in one period interval 0  n  4 equals to x½n. And one period of the shifted periodic signal for 0  n  4 can be obtained by rotate right operation as ~xop ½n  2 ¼ RRðx½n; 2Þ which can be calculated in two steps as follows ~xop ½n  1 ¼ RRðx½n; 1Þ ¼ ½ 1 0:5

0:5 0:5

1



~xop ½n  2 ¼ RR ~xop ½n  1; 1 ¼ ½1

1 0:5

0:5 0:5 :

As a result x½ðn  2Þ5  is found as  x ð n  2Þ 5 ¼ ½ 1

1 0:5

0:5

0:5 :

Example 3.32 If x1 ½n ¼ ½ 1 1 1 0:5  and x2 ½n ¼ ½ 1 find 4-point circular convolution of x1 ½n and x2 ½n. That is,

1 1

1 ,

Solution 3.32 Method 1: N-point circular convolution of x1 ½n and x2 ½n can be calculated using ð3:141Þ

Let N ¼ 4 we get

expanding the right hand side of (3.141) for

   y½n ¼ x1 ½0x2 ðnÞ4 þ x1 ½1x2 ðn  1Þ4 þ x1 ½2x2 ðn  2Þ4  þ x1 ½3x2 ðn  3Þ4

ð3:142Þ

    where the signals x2 ðnÞ4 , x2 ðn  1Þ4 , x2 ðn  2Þ4 , and x2 ðn  3Þ4 can be calculated as

194

3

Discrete Fourier Transform

  x2 ðnÞ4 ¼ x2 ½n ! x2 ðnÞ4 ¼ ½ 1 1 1 1    x2 ðn  1Þ4 ¼ RRðx2 ½n; 1Þ ! x2 ðn  1Þ4 ¼ ½ 1 1 1 1    x2 ðn  2Þ4 ¼ RRðx2 ½n; 2Þ ! x2 ðn  2Þ4 ¼ ½ 1 1 1 1    x2 ðn  3Þ4 ¼ RRðx2 ½n; 3Þ ! x2 ðn  3Þ4 ¼ ½ 1 1 1 1 :

ð3:143Þ

Substituting the calculated values in (3.143) into (3.142), we get y½n ¼ ð1Þ  ½ 1 1 1 1  þ ð1Þ  ½ 1 1 1 1  þ ð1Þ  ½ 1 1 1 1  þ ð0:5Þ  ½ 1 1 1 1  which is simplified as y½n ¼ ½ 3:5 0:5

3:5 0:5 :

Method 2: N-point circular convolution of x1 ½n and x2 ½n can be calculated as y ½ n ¼

N1 X

 x1 ½mx2 ðn  mÞN :

ð3:144Þ

m¼0

Evaluating the right hand side of (3.144) for the n values in the range 0  n  N  1, we get the equation set y½0 ¼ y½1 ¼

N1 P m¼0 N1 P m¼0

y ½ N  1 ¼

N1 P m¼0

 x1 ½mx2 ð0  mÞN  x1 ½mx2 ð1  mÞN .. .

 x1 ½mx2 ðN  1  mÞN :

For N ¼ 4 equation set (3.145) becomes as

ð3:145Þ

3.5 Discrete Fourier Transform

195

y½0 ¼

3 X

 x1 ½mx2 ð0  mÞ4

m¼0

y½1 ¼

3 X

 x1 ½mx2 ð1  mÞ4

m¼0

y½2 ¼

3 X



x1 ½mx2 ð2  mÞ4

ð3:146Þ



m¼0

y½3 ¼

3 X

 x1 ½mx2 ð3  mÞ4

m¼0

    where the signals x2 ðmÞ4 , x2 ð1  mÞ4 , x2 ð2  mÞ4 , and x2 ð3  mÞ4 are calculated as   x2 ðmÞ4 ¼ RI ðx2 ½mÞ ! x2 ðmÞ4 ¼ ½ 1 1 1 1   

 x2 ð1  mÞ4 ¼ RR x2 ðmÞ4 ; 1 ! x2 ð1  mÞ4 ¼ ½ 1 1 1 1   

 x2 ð2  mÞ4 ¼ RR x2 ð1  mÞ4 ; 1 ! x2 ð2  mÞ4 ¼ ½ 1 1 1 1   

 x2 ð3  mÞ4 ¼ RR x2 ð2  mÞ4 ; 1 ! x2 ð3  mÞ4 ¼ ½ 1 1 1 1 : Now consider the summation term y ½ 0 ¼

3 X

 x1 ½mx2 ð0  mÞ4 :

ð3:147Þ

m¼0

 Let w½m ¼ x2 ðmÞ4 i.e., w½m ¼ ½ 1 we obtain

1 1

1 ; then expanding (3.147),

y½0 ¼ x1 ½0w½0 þ x1 ½1w½1 þ x1 ½2w½2 þ x1 ½3w½3 which is nothing but dot product of two vectors x1 ½n and w½n, that is y½0 ¼ ½ x1 ½0 x1 ½1

x1 ½2 x1 ½3   ½ w½0 w½1

which can also be written as y½0 ¼ x1 ½m  w½m or  y½0 ¼ x1 ½m  x2 ðmÞ4 : Then

w½2 w½3 

196

3

Discrete Fourier Transform

y½0 ¼ ð1Þ  ð1Þ þ ð1Þ  ð1Þ þ ð1Þ  ð1Þ þ ð0:5Þ  ð1Þ y½0 ¼ 3:5: In a similar manner,  y½1 ¼ x1 ½m  x2 ð1  mÞ4 y½1 ¼ ð1Þ  ð1Þ þ ð1Þ  ð1Þ þ ð1Þ  ð1Þ þ ð0:5Þ  ð1Þ y½1 ¼ 0:5  y½2 ¼ x1 ½m  x2 ð2  mÞ4 y½2 ¼ ð1Þ  ð1Þ þ ð1Þ  ð1Þ þ ð1Þ  ð1Þ þ ð0:5Þ  ð1Þ y½2 ¼ 3:5  y½3 ¼ x1 ½m  x2 ð3  mÞ4 y½3 ¼ ð1Þ  ð1Þ þ ð1Þ  ð1Þ þ ð1Þ  ð1Þ þ ð0:5Þ  ð1Þ y½3 ¼ 0:5 As a result; y½n ¼ ½ 3:5 0:5

3:5 0:5 :

and the length of x1 ½n or x2 ½n is shorter Note: If than N then the shorter sequence is padded by zeros so that its length equals to N. If both sequences are shorter than N samples then both sequences are padded by zeros so that their lengths equal to N. Example 3.33 If x1 ½n ¼ ½ 1 1 1 0:5  and x2 ½n ¼ ½ 1 6-point circular convolution of x1 ½n and x2 ½n. That is,

1

2 , find

Solution 3.33 The lengths of the sequences x1 ½n and x2 ½n are 4 and 3 respectively. Both sequences should be padded by zeros so that their lengths equals to 6. That is, x1 ½n ¼ ½ 1

1

1

0:5 0

0

x 2 ½ n ¼ ½ 1

1 2

0

0 0 :

Then circular convolution operations can be performed as in Example 3.32. Matrix Representation of Circular Convolution Example 3.34 If x1 ½n ¼ ½ x1 ½0 x1 ½1 x1 ½2  x2 ½n ¼ ½ x2 ½0 x2 ½1 x2 ½2  Express 3-point circular convolution of x1 ½n and x2 ½n as matrix multiplication.

3.5 Discrete Fourier Transform

197

Solution 3.34 Expanding the expression y½n ¼

N 1 X

 x1 ½mx2 ðn  mÞN

m¼0

for N ¼ 3, we get    y½n ¼ x1 ½0x2 ðnÞN þ x1 ½1x2 ðn  1ÞN þ x1 ½2x2 ðn  2ÞN which is calculated as y½n ¼ x1 ½0 x2 ½0 x2 ½1 x2 ½2 þ x1 ½1½ x2 ½2 þ x1 ½2½ x2 ½1 x2 ½2 x2 ½0 :

x2 ½0 x2 ½1 

The expression in (3.48) can be written using matrix multiplication as 0

1 y½0 @ y½1 A ¼ ð x1 ½0 y½2 Example 3.35 If x½n ¼ ½1

0

x 2 ½ 0 x 2 ½ 2 x1 ½1 x1 ½2 Þ  @ x2 ½1 x2 ½0 x 2 ½ 2 x 2 ½ 1 0 |{z} 1

1 2

1 x 2 ½ 1 x 2 ½ 2 A : x 2 ½ 0

1, find

n¼0

Solution 3.35 Since the index n ¼ 0 is not at the first element in x½n, it is easier to calculate the circular convolution using the first method we introduced. That is expanding 3 X  x1 ½mx1 ðn  mÞN y½n ¼ m¼2

for m values, we obtain    y½n ¼ x1 ½2x1 ðn þ 2Þ6 þ x1 ½1x1 ðn þ 1Þ6 þ x1 ½0x1 ðnÞ6    þ x1 ½1x1 ðn  1Þ6 þ x1 ½2x1 ðn  2Þ6 þ x1 ½3x1 ðn  3Þ6 and placing the n values in the range 2  n  3 for y½n, we can find the 6-point circular convolution result.

198

3

Discrete Fourier Transform

Exercise: Prove the following property  DFT 2p x ðn  mÞN $ ej N km X ½k : The Relationship between Circular and Linear Convolution: x1 ½n ! Aperiodic signal with length L x2 ½n ! Aperiodic signal with length P Linear convolution of x1 ½n and x2 ½n is calculated using ylc ½n ¼

1 X

x1 ½mx2 ½n  m:

m¼1

The length of ylc ½n is L þ P  1. N-point circular convolution of x1 ½n and x2 ½n is The relationship between ylc ½n and ycc ½n is given as 8 1 < P y ½n  rN  0  n  N  1 ycc ½n ¼ r¼1 lc : 0 otherwise: If N  L þ P  1 then the circular convolution and linear convolution results are the same, i.e., ylc ½n ¼ ycc ½n.

3.6

Practical Calculation of the Linear Convolution

Overlap Add and Overlap Save Methods For practical communication systems, the input signal may not be of finite duration. It may be of infinite duration or may be a very long sequence, such as TV signal, video or speech signal. The input signal x½n is usually passed through a filter with an impulse response h½n. Filtering operation is nothing but the convolution of the input signal with the impulse response of the filter, the filter output is y½n ¼ x½n h½n: If the input signal is very long, then convolution operation takes too much time, or sometimes it may not still be possible to evaluate the convolution result.

3.6 Practical Calculation of the Linear Convolution

199

To overcome this issue, two approaches are followed to evaluate the convolution of a very long input and a short impulse response sequences. These methods are called overlap-add and overlap-save. Let’s first explain the overlap-add method.

3.6.1

Evaluation of Convolution Using Overlap-Add Method

Let x½n be the input signal with length N and h½n be the filter response with length P such that N [ P. The overlap-add method to evaluate x½n h½n consists of the following steps: (1) Divide the input sequence to frames such that each frame has length L. Let’s denote the frames by x0 ½n; x1 ½n; x2 ½n   

0nL  1

(2) Evaluate the convolution of each frame with h½n, i.e., evaluate yk ½n ¼ xk ½n h½n k ¼ 0; 1; 2; . . . (3) Calculate the convolution result as y ½ n ¼

1 X

yk ½n  Lk:

k¼0

Let’s explain overlap-add method with an example. Example 3.36 If x½n ¼ ½ 1 1 0 1 1 0 1 1 1 h½n ¼ ½ 1 1 , find x½n h½n using overlap-add method.

1

0 1  and

Solution 3.36 (1) In step 1, we divide the input sequence into frames of length L. The length of the impulse response h½n is P ¼ 2. The length of the frames depends on our choice. Let’s choose the length of the frames as L ¼ 3 and divide the sequence x½n into frames as shown in (3.148) x½n ¼ ½1 1 0 |fflfflfflfflfflffl{zfflfflfflfflfflffl} x0 ½n

1 1 0 |fflfflfflfflfflffl{zfflfflfflfflfflffl} x1 ½n

1 1 1 |fflfflfflfflfflffl{zfflfflfflfflfflffl} x½n

1|fflfflfflfflfflffl{zfflfflfflfflfflffl} 0 1

ð3:148Þ

x3 ½n

If the last frame had a length smaller than 3, then we would pad it by zeros until its length equals to 3. The divided frames are

200

3

x0 ½n ¼ ½ 1 x 2 ½ n ¼ ½ 1

1 0  x 1 ½ n ¼ ½ 1

1 0 

1  x3 ½n ¼ ½ 1

1

Discrete Fourier Transform

0

ð3:149Þ

1 :

(2) In step 2, we take the convolution of each frame in (3.149) with impulse response h½n. Let’s first calculate the convolution of x0 ½n and h½n, i.e., calculate y0 ½n ¼ x0 ½n h½n which is written as 1 X

y 0 ½ n ¼

h½kx½n  k

ð3:150Þ

k¼1

When (3.150) is expanded for n ¼ 0; 1; 2; 3; we obtain y0 ½n ¼ x0 ½n h½n ! y0 ½n ¼ ½ 1 y1 ½n ¼ x1 ½n h½n ! y0 ½n ¼ ½ 1 y2 ½n ¼ x2 ½n h½n ! y0 ½n ¼ ½ 1

1 0 ½1

1  ! y0 ½n ¼ ½ 1 2

1 0 

1 0  ½ 1 1  ! y1 ½n ¼ ½ 1 2 1 0  1 1  ½ 1 1  ! y2 ½n ¼ ½ 1 2 1 0 

y3 ½n ¼ x3 ½n h½n ! y0 ½n ¼ ½ 1 0

1  ½ 1

1  ! y3 ½n ¼ ½ 1 1

1 1 :

ð3:151Þ (3) In this step, using the results of (3.151) in y ½ n ¼

1 X

yk ½n  Lk

k¼0

Fig. 3.29 Shifting of y1 ½n

y1[n] 1

2

1

0

0

1

2

3

n

y1[n 3] 0

0

0

1

2

1

0

0

1

2

3

4

5

6

n

3.6 Practical Calculation of the Linear Convolution

201

for L ¼ 3, we obtain y½n ¼

3 X

yk ½n  k3

k¼0

which is expanded as y½n ¼ y0 ½n þ y1 ½n  3 þ y2 ½n  6 þ y3 ½n  9:

ð3:152Þ

The signal y1 ½n  3 in (3.152) is obtained by shifting the amplitudes of y1 ½n to the right by 3 units. When amplitudes are shifted to the right, zero amplitude values are inserted into the old positions. This means that y1 ½n  3 can be obtained by padding 3 zeros to the beginning of y1 ½n. This operation is illustrated in Fig. 3.29. Thus, the shifted signals together with y0 ½n can be written as

x[n]

0

L 1 L

n

2L

Fig. 3.30 Dividing x½n into frames

w0 [ n] 0

0

n

L 1

w1[ n] L

0 Fig. 3.31 Divided frames of x½n are shown separately

2L 1

2L

n

w2 [n] n

202

3

Discrete Fourier Transform

x0 [n]

0

n

L 1

x1[n]

0

n

L 1 x2 [ n ]

n

0

L 1

Fig. 3.32 Divided frames of x½n start at n ¼ 0

x0 [n] 0

n

L 1

0

x1[n L]

L

n

2L 1 x2 [ n 2 L]

0

n

2L

Fig. 3.33 Frames starting at n ¼ 0 are shifted by multiples of L

y0 ½n ¼ ½|{z} 1 0 y1 ½n  3 ¼ ½|{z}

1

2

n¼0

0

0 0

1

2 1

0 0

0

0

0 1

0 0

0

0

0

n¼0

y2 ½n  6 ¼ ½|{z} 0 0 y3 ½n  9 ¼ ½|{z}

0

n¼0

0

0

0

0 1

ð3:153Þ

2 1 1

1

1:

n¼0

When the shifted signals in (3.153) are summed, we obtain the convolution result as

3.6 Practical Calculation of the Linear Convolution

y½n ¼ ½ 1

2

1

1

2

1

1

203

0

2

2

1 1

1 :

Now let’s see the mathematical derivation of the overlap-add method. Assume that the digital sequence x½n is divided into frames as shown in Fig. 3.30. And the frames are separately shown in Fig. 3.31. Let’s make the starting index of every frame be equal to n ¼ 0. This is shown in Fig. 3.32. We can obtain the digital signal x½n by shifting and summing the frames that starts at n ¼ 0 as shown in Fig. 3.33. This operation is mathematically written as x ½ n ¼

1 X

xk ½n  Lk:

k¼0

Then the convolution of x½n and h½n can be written as y ½ n ¼ h½ n x ½ n 1 X ¼ h½ n xk ½n  Lk:

ð3:154Þ

k¼0

When the summation term in (3.154) is expanded, we get y½n ¼ h½n ðx0 ½n þ x1 ½n  L þ x2 ½n  2L þ   Þ:

ð3:155Þ

And for linear time invariant systems if y1 ½n ¼ h½n x1 ½n then y1 ½n  L ¼ h½n x1 ½n  L: Using a similar approach for the other convolutional expressions appearing in (3.155), we get y½n ¼ y0 ½n þ y1 ½n  L þ y2 ½n  2L þ    where y0 ½n ¼ h½n x0 ½n;

y1 ½n ¼ h½n x1 ½n;

y2 ½n ¼ h½n x2 ½n:

204

3

Discrete Fourier Transform

As a result; y ½ n ¼

1 X

yk ½n  Lk:

k¼0

3.6.2

Overlap-Save Method

Assume that the impulse response h½n has length P. The convolution of x½n and h½n using overlap-save method is achieved via the following steps. (1) Pad the front of x½n by P  1 zeros. (2) Divide x½n into frames of length L such that the successor frame overlaps with the predecessor frame with P  1 points. (3) Let xk ½n be a frame, calculate the L point circular convolution of xk ½n and h½n, i.e., calculate yk ½n ¼ xk ½mðLÞh½n: (4) Discard the first P  1 points of yk ½n. (5) Concatenate yk ½n and obtain y½n, i.e., y½n ¼ ½y0 ½ny1 ½n   : Let’s explain overlap-save method with an example. Example 3.37 Using h½n and x½n given below, find the convolution of h½n and x½n using overlap-save method. h½n ¼ ½ 1 x ½ n ¼ ½ 1 0

1

1 1

1

1

1 1 0

0

1

1 1 

Take frame length as L ¼ 4. Solution 3.37 The length of the impulse response h½n is 3, i.e., P ¼ 3. And frame length is L ¼ 4 which is given the question, otherwise we can choose it according to our will. Let’s follow the steps of the overlap-save method for the calculation of convolution of h½n and x½n. (1) Add P  1 ¼ 3  1 ! 2 zeros to the beginning of x½n. This is shown in

3.6 Practical Calculation of the Linear Convolution

x½n ¼ ½

0 0 |ffl{zffl} P  1 zeros

1 0

1

205

1 1

1

1

0 0

1

 1

1

are added to the beginning of x½n (2) Divide x½n into frames such that frames overlap by P  1 ¼ 2 samples. This operation is illustrated in

where we padded the last divided frame by 2 zeros such that its length equals 4. The divided frames are separately written as x 0 ½ n ¼ ½ 0 0 x3 ½n ¼ ½ 1 1

0

1 1

x 1 ½ n ¼ ½ 1

0 1

0  x 4 ½ n ¼ ½ 1 0

0

1

x2 ½n ¼ ½ 1

1

x 5 ½ n ¼ ½ 0

1

1

1 1

1  1 

(3) In step 3 we calculate the L ¼ 4-point circular convolution of each frame with h½n, i.e., we calculate y0 ½n ¼ h½nð4Þx0 ½n y1 ½n ¼ h½nð4Þx1 ½n y2 ½n ¼ h½nð4Þx2 ½n y3 ½n ¼ h½nð4Þx3 ½n y4 ½n ¼ h½nð4Þx4 ½n y5 ½n ¼ h½nð4Þx5 ½n : As a reminder we below provide the 4-points circular convolution of x0 ½n and h½n. N-point circular convolution of x½n and h½n is given as y0 ½ n ¼

N 1 X

x½kh½ðn  kÞN :

ð3:156Þ

k¼0

For N ¼ 4 when (3.156) is expanded, we obtain     y0 ½n ¼ x½0h ðnÞ4 þ x½1h ðn  1Þ4 þ x½2h ðn  2Þ4 þ x½3h ðn  3Þ4 : Since N ¼ 4 we pad h½n by zeros such that its length equals N ¼ 4 and h½n becomes as

206

3

h½n ¼ ½ 1

1 1

Discrete Fourier Transform

0 :

 Noting that h ðn  n0 Þ4 is obtained rotating h½n to the right by n0 units, we get the following expression for y0 ½n     y 0 ½ n ¼ 0  h ð nÞ 4 þ 0  h ð n  1Þ 4 þ 1  h ð n  2Þ 4 þ 0  h ð n  3Þ 4 which leads to y 0 ½ n ¼ ½ 1 0

1

1 :

4-point circular convolution of each frame with h½n is given in (3.157). y0 ½n ¼ ½ 1

0

1 1  y1 ½n ¼ ½ 1

y3 ½n ¼ ½ 0 2 1 0  y4 ½n ¼ ½ 0 y6 ½n ¼ ½ 1 2 0 1 

2 2 

2 2

1

y 2 ½ n ¼ ½ 1

1  y5 ½n ¼ ½ 0

3 2 0

1

3

3 ð3:157Þ

(4) In step-4, we discard the first P  1 ¼ 2 samples from the beginning of each yk ½n; k ¼ 0; 1; 2; 3; 4: This operation is illustrated in   1 0 1 1 y0 ½n ¼ |fflffl{zfflffl} ! y0 ½n ¼ ½ 1 1   omit  1 2 2 2 y1 ½n ¼ |fflfflfflffl{zfflfflfflffl} ! y1 ½n ¼ ½ 2 2   omit  1 3 1 3 y2 ½n ¼ |fflfflfflffl{zfflfflfflffl} ! y2 ½n ¼ ½ 1 3   omit  0 2 1 0 y3 ½n ¼ |fflffl{zfflffl} ! y3 ½n ¼ ½ 1 0   omit  0 2 1 1 y4 ½n ¼ |fflffl{zfflffl} ! y4 ½n ¼ ½ 1 1   omit  0 2 0 3 y5 ½n ¼ |fflfflfflffl{zfflfflfflffl} ! y 5 ½ n ¼ ½ 0 3   omit  1 2 0 1 y6 ½n ¼ |fflfflfflffl{zfflfflfflffl} ! y5 ½n ¼ ½ 0 1 : omit

(5) Finally in the last step, we concatenate the truncated sequences to find the convolution result, i.e.,

3.6 Practical Calculation of the Linear Convolution

207

y½n ¼ ½y0 ½ny1 ½ny2 ½ny3 ½ny4 ½ny5 ½n which leads to y½n ¼ ½ 1

1

2 2

Exercise: If x½n ¼ ½ 1

1 3

1 1

1

1 1

0 1

1

0 3

0

1 :

1 1 111111111 and

h½n ¼ ½ 1 1 1 , calculate x½n h½n (a) Using overlap-add method. (b) Using overlap-save method.

3.7 3.7.1

Computation of the Discrete Fourier Transform Fast Fourier Transform (FFT) Algorithms

There are two types of Fast Fourier transform algorithm. These are: (1) Decimation in time FFT algorithm. (2) Decimation in frequency FFT algorithm. Let’s first explain decimation in time FFT algorithm then decimation in frequency FFT algorithm.

3.7.2

Decimation in Time FFT Algorithm

Before starting to the derivation of the algorithm, let’s consider some motivating examples. The DFT formula is X ½k  ¼

N 1 X

x½nejk N n 2p

n¼0

where k takes values in the range 0; 1; . . .; N  1, i.e., if N ¼ 4, then the range of k is 0; 1; 2; 3.

208

3

Discrete Fourier Transform

Example 3.38 If ekN is defined as ekN ¼ ejk N k 2 Z, write ek4 for k ¼ 0; 1; 2; 3 as a vector.  2p 2p 2p 2p Solution 3.38 ek4 ¼ ej0 4 ej1 4 ej2 4 ej3 4 which can be simplified as 2p

ek4 ¼ ½ 1

j 1

j

Exercise: Write ek8 for k ¼ 0; 1; . . .; 7 as a vector. Example 3.39 Given x½n ¼ ½ a

b  find 2-point DFT of x½n.

Solution 3.39 Using the formula X ½k  ¼

N 1 X

x½nejk N n ; 2p

k ¼ 0; 1; . . .; N  1

n¼0

for N ¼ 2, we get X ½k  ¼

1 X

x½nejk 2 n ; 2p

k ¼ 0; 1

ð3:158Þ

n¼0

When (3.158) is expanded for k ¼ 0 and k ¼ 1, we get X ½0 ¼ x½0 þ x½1 X ½1 ¼ x½0 þ x½1ejp ! X ½1 ¼ x½0  x½1: Then 2-point DFT of x½n ¼ ½ a

b  is

X ½k ¼ ½ a þ b a  b : Example 3.40 If x½n ¼ ½ 3

2 , find 2-point DFT of x½n.

Solution 3.40 Using X ½k  ¼ ½ a þ b ½ 3 2  as

a  b , we find the 2-point DFT of x½n ¼

X ½k  ¼ ½ 1 Example 3.41 If x½n ¼ ½ a

5 :

b , find X½k for k ¼ 0; 1; 2; 3:

Solution 3.41 Expanding the formula

3.7 Computation of the Discrete Fourier Transform

X ½k  ¼

N1 X

209

x½nejk N n 2p

n¼0

for N ¼ 2 and k ¼ 0; 1; 2; 3; we obtain X ½ 0 ¼

1 X

x½nej0 2 n 2p

n¼0

X ½ 1 ¼

1 X

x½nej1 2 n 2p

n¼0

X ½ 2 ¼

1 X

x½nej2 2 n 2p

n¼0

X ½ 3 ¼

1 X

x½nej3 2 n : 2p

n¼0

If we look at the exponential terms in X½0 and X½2, we see that ej0 2 n ¼ ej2 2 n this means that X ½2 ¼ X½0 2p

In a similar manner; we find that X ½3 ¼ X½1 Then X½k for k ¼ 0; 1; 2; 3; happens to be " X ½k  ¼

X ½0 X ½1

X ½ 2 |{z} ¼X½0

X ½3 |{z}

#

¼X½1

That is X ½k ¼ ½ X ½0

X ½ 1 X ½ 0 X ½ 1 

And using our previous example results, we can write X½k as X ½k  ¼ ½ a þ b a  b

aþb

a  b

Example 3.42 Calculate X½k for x½n ¼ ½ 1

3

2

1 

using the DFT formula but take k range as 0; 1; . . .; 7 instead of 0; 1; . . .; 3.

2p

210

3

Discrete Fourier Transform

Solution 3.42 Using the DFT formula the DFT coefficients for k ¼ 0; 1; . . .; 7 can be calculated as X ½ 0 ¼ X ½ 1 ¼ X ½ 2 ¼ X ½ 3 ¼ X ½ 4 ¼ X ½ 5 ¼ X ½ 6 ¼ X ½ 7 ¼

3 P n¼0 3 P n¼0 3 P n¼0 3 P n¼0 3 P n¼0 3 P n¼0 3 P n¼0 3 P

x½nej0 4 n 2p

x½nej1 4 n 2p

x½nej2 4 n 2p

x½nej3 4 n 2p

ð3:159Þ x½ne

j42p 4n

x½nej5 4 n 2p

x½nej6 4 n 2p

x½nej7 4 n : 2p

n¼0

If we inspect the exponential terms in X½0 and X½4 in the equation set (3.159), 2p 2p we see that ej0 4 n ¼ ej4 4 n this means that X ½4 ¼ X ½0: In a similar manner; we have X ½5 ¼ X ½1 X ½6 ¼ X ½2 X ½7 ¼ X ½3: If we calculate X½k for k ¼ 0; 1; . . .; 3; we get X ½k  ¼ ½ X ½0

X ½1 X ½2 X ½3 :

And on the other hand if we calculate X½k for k ¼ 0; 1; . . .; 7; we get " X ½k  ¼ That is

X ½0

X ½ 1 X ½ 2 X ½ 3

X ½ 4 |{z} ¼X½0

X ½5 |{z} ¼X½1

X ½ 6 |{z} ¼X½2

# X ½7 |{z} : ¼X½3

3.7 Computation of the Discrete Fourier Transform

X ½ k  ¼ ½ X ½ 0 X ½ 1

X ½2

211

X ½ 3 X ½ 0 X ½ 1

X ½2

X ½3 :

Using (3.159), X½k for k ¼ 0; 1; 2; 3 can be calculated as X ½k  ¼ ½ 5

1  j4

1 1 þ j4 

1 þ j4

5

and for k ¼ 0; 1; . . .; 7; it equals to X ½k  ¼ ½ 5

1  j4

1

1  j4

1

1 þ j4 :

In fact, the results of these examples are nothing but the main motivation for the derivation of the fast Fourier transform algorithm. Now let’s start the derivation of the fast Fourier transform algorithm. Fast Fourier Transform Algorithm Derivation We consider the DFT formula X ½k  ¼

N 1 X

x½nejk N n : 2p

ð3:160Þ

n¼0

Let’s denote the exponential function ej N in (3.160) by eN , i.e., eN ¼ ej N , and the function eN has the following properties. 2p

2p

(1) e2N ¼ eN=2 This property comes from the definition directly, i.e., e2N=2 ¼ ej2 N

2p

which can be written as e2N=2 ¼ ejN=2 ¼ eN=2 ! e2N ¼ eN=2 : 2p

(2) eNN ¼ 1 or more in general emN N ¼ 1; m 2 Z Again starting by the definition, we have jm N jm2p eN ¼ ej N ! emN ! emN ! emN N ¼e N ¼ e N ¼ 1: 2p

ðm þ N Þ

(3) eN

ðmÞ

¼ eN

Using property-2 we obtain

2pN

212

3 ðm þ N Þ

eN

ðmÞ

ðN Þ

ðm þ N Þ

¼ eN eN ! eN |{z}

Discrete Fourier Transform

ðmÞ

¼ eN :

¼1

ðmÞ

This means that f ðmÞ ¼ eN is a periodic function, and its period equals to N, i.e., f ðmÞ ¼ f ðm þ NÞ. Let’s now derive the decimation in time FFT algorithm. We first write the DFT formula in terms of the defined function eN as X ½k  ¼

N 1 X

ðknÞ

x½neN ;

k ¼ 0; 1; . . .; N  1

ð3:161Þ

n¼0

which can be partitioned for even and odd n values as X ½k  ¼

N=21 X

ð2knÞ

x½2neN

N=21 X

þ

n¼0

ð2n þ 1Þk

x½2n þ 1eN

ð3:162Þ

n¼0

where the first term on the right side using the property e2N ¼ eN=2 can be written as N=21 X

ð2nk Þ

x½2neN

!

N=21 X

n¼0

x½2nðe2N Þnk !

n¼0

N=21 X

x½2nðeN=2 Þnk

ð3:163Þ

n¼0

and the similarly the second term on the right side of (3.162) using the property e2N ¼ eN=2 can be written as N=21 X

ð2n þ 1Þk

x½2n þ 1eN

!

N=21 X

n¼0

!

k x½2n þ 1e2nk N eN

n¼0

ekN

N=21 X

x½2n þ 1e2nk N

!

ekN

N=21 X

n¼0

ð3:164Þ x½2n þ 1enk N=2

n¼0

Then using the results (3.163) and (3.164), the DFT formula in (3.161) can be written as X ½k  ¼

N=21 X

k x½2nenk N=2 þ eN

n¼0

N=21 X

x½2n þ 1enk N=2

n¼0

|fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl}

|fflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflffl}

G½k

H½k

k ¼ 0; 1; . . .; N  1

3.7 Computation of the Discrete Fourier Transform

213

where the terms G½k and H½k are periodic with period N=2. Since G½k and H½k are calculated for k ¼ 0; 1; . . .N  1 in X½k then G½k and H½k have repeated values for k ¼ 0; 1; . . .N  1 as shown in 2

3 g0 g1 g2    |fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl} 7 The secondN=2 5 samples

2

3 h0 h1 h2    |fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl} 7 The second N=2 5: samples

g0 g1 g2    6 |fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl} G½k ¼ 4 The first N=2 samples h0 h1 h2    6 |fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl} H ½k  ¼ 4 The first N=2 samples

And G½k k ¼ 0; 1; . . .; N=2  1 is the N=2 point DFT of the even numbered samples of x½n, and H½k k ¼ 0; 1; . . .; N=2  1 is the N=2 point DFT of the odd numbered samples of x½n. Hence for the computation of G½k and H½k the k index range is first taken as k ¼ 0; 1; . . .; N=2  1. And G½k and H½k are calculated for k ¼ 0; 1; . . .; N=2  1. Let’s denote the calculation results as  G½k  ¼ g0

g1

   gN=21



 H ½ k  ¼ h0

h1

   hN=21



k ¼ 0; 1; . . .; N=2  1

Then G½k and H½k values for k ¼ 0; 1; . . .; N  1 are obtained using G ½k  ¼ H ½k  ¼

 

g0

g1



gN=21

g0

g1

   gN=21

h0

h1



hN=21

h0

h1

   hN=21



and they are combined in X½k via X ½k ¼ G½k þ wkN H ½k

k ¼ 0; 1; . . .; N  1:

The partition performed for X ½k  can be done for G½k  and H½k also. The calculation of G½k  can be written as G½k ¼ G1 ½k þ wkN G2 ½k  k ¼ 0; 1; . . .; N=2  1 where G1 ½k is the N=4 point DFT of the even numbered samples of x½2n and G2 ½k is the N=4 point DFT of the odd numbered samples of x½2n. And the calculation of H½k can be written as H ½k ¼ H1 ½k þ wkN H2 ½k

k ¼ 0; 1; . . .; N=2  1

where H1 ½k is the N=4 point DFT of the even numbered samples of x½2n þ 1 and H2 ½k is the N=4 point DFT of the odd numbered samples of x½2n þ 1.

214

3

Discrete Fourier Transform

This procedure can be carried out until we calculate 2-point DFT of the sequences obtained from x½n. Example 3.43 If x½n ¼ ½ a

b  find 2-point DFT of x½n.

Solution 3.43 Using the formula X ½k  ¼

N 1 X

x½nejk N n ; 2p

k ¼ 0; 1; . . .; N  1

ð3:165Þ

n¼0

for N ¼ 2, we get X ½k  ¼

1 X

x½nejk 2 n ; 2p

k ¼ 0; 1:

ð3:166Þ

n¼0

When (3.166) is expanded for k ¼ 0 and k ¼ 1, we obtain X ½0 ¼ x½0 þ x½1 X ½1 ¼ x½0 þ x½1ejp ! X ½1 ¼ x½0  x½1 which can be expressed in a more compact way as X ½k ¼ ½ a þ b a  b : Example 3.44 If x½n ¼ ½ 1

ð3:167Þ

4 , find 2-point DFT of x½n.

Solution 3.44 X ½0 ¼ 1 þ 4 ! X ½0 ¼ 3 Example 3.45 If x½n ¼ ½ 1 1 1 mation in time FFT algorithm.

X ½1 ¼ 1  4 ! X ½1 ¼ 5:

2 , find 4-point DFT of x½n using deci-

Solution-3.45: First we determine the even and odd numbered elements of x½n as in 2

"

x½n ¼ 4 |{z} 1 #

z}|{ 1

1 |{z}

3 " z}|{ 2 5

#

where down-arrows indicate even numbered samples and up-arrows show odd numbered samples. And the even and odd numbered samples can be grouped into separate vectors as x e ½ n ¼ ½ 1

1  xo ½n ¼ ½ 1

2 :

The 2-point DFT of xe ½n and xo ½n are calculated using DFT formula as

3.7 Computation of the Discrete Fourier Transform

Xe ½0 ¼ 1  1 ! Xe ½0 ¼ 0 Xo ½0 ¼ 1 þ 2 ! Xe ½0 ¼ 3

215

Xe ½1 ¼ 1  ð1Þ ! Xe ½1 ¼ 2 Xo ½1 ¼ 1  2 ! Xe ½1 ¼ 1:

Hence, for Xe ½k and Xo ½k k ¼ 0; 1; we have Xe ½k  ¼ ½ 0

2  X o ½k  ¼ ½ 3

1 :

ð3:168Þ

DFT of x½n can be written in terms of DFT of its even and odd samples as X ½k  ¼ Xe ½k  þ wkN Xo ½k

k ¼ 0; 1; . . .; N  1:

ð3:169Þ

k ¼ 0; 1; . . .; 4  1

ð3:170Þ

For N ¼ 4 Eq. (3.169) is written as X ½k ¼ Xe ½k  þ wk4 Xo ½k where wk4 ¼ ejk 4 : 2p

And for N ¼ 4 the vectors Xe ½k, Xo ½k and wk4 for k ¼ 0; 1; 2; 3 can be calculated as Xe ½k ¼ ½ 0 2 0 2  X0 ½k ¼ ½ 3 1  2p 2p 2p 2p wk4 ¼ ej0 4 ej1 4 ej2 4 ej3 4 :

3 1 

ð3:171Þ

And simplifying wk4 , we get wk4 ¼ ½ 1 j

1

j :

Finally the vector X½k is obtained using (3.170) as in X ½k  ¼ ½ 0

2

0 2  þ ½  j 1

j ½3

1

where the vector product term ½ 1 j 1

j  ½ 3 1

3

1 

is calculated as ½1  3 Then X ½k becomes as

ðjÞ  ð1Þ

ð1Þ  3

j  ð1Þ :

3 1 

216

3

X ½k  ¼ ½ 0 þ 1  3 2 þ ðjÞ  ð1Þ

Discrete Fourier Transform

0 þ ð1Þ  3 2 þ j  ð1Þ 

which has the final form X ½k  ¼ ½ 3

2 þ j 3

Example 3.46 If x½n ¼ ½ 1 1 1 2 1 using decimation in time FFT algorithm.

2  j : 1

3

2 , find 8-point DFT of x½n

Solution 3.46 First, we divide the sequence x½n to its even and odd numbered elements as in 2

"

x½n ¼ 4 |{z} 1

z}|{ 1

#

"

1 |{z}

"

z}|{ 2

1 |{z}

#

z}|{ 3

1 |{z}

#

3 " z}|{ 2 5

#

where down-arrows indicate even indexed samples and up-arrow shows odd indexed samples. And the even and odd indexed samples can be grouped into separate vectors as xe ½n ¼ ½ 1

1

1 1 

x o ½ n ¼ ½ 1

2

3 2 :

Four-point DFT of xe ½n and xo ½n can be calculated as in the previous example as Xe ½k  ¼ ½ 0 0

4

0

Xo ½k ¼ ½ 8

2

0 2  k ¼ 0; 1; . . .; 4:

ð3:172Þ

Then 8-point DFT of x½n is calculated through X ½k ¼ Xe ½k þ wk8 Xo ½k k ¼ 0; 1; . . .; 7 where wk8 ¼ ejk 8 . And the vectors Xe ½k, Xo ½k , wk8 for k ¼ 0; 1; . . .; 7 with the help of (3.172) can be written as 2p

Xe ½k ¼ ½ 0

0 4

0

0

0 4

Xo ½k ¼ ½ 8 2 0 2 8 2 0 2   2p 2p 2p 2p 2p 2p wk8 ¼ ej0 8 ej1 8 ej2 8 ej3 8 ej4 8 ej5 8

0

ej6 8

2p

And combining the vectors in (3.173) using X ½k ¼ Xe ½k þ wk8 Xo ½k k ¼ 0; 1; . . .; 7

ej7 8

2p



ð3:173Þ

3.7 Computation of the Discrete Fourier Transform

217

we obtain the 8-point DFT of x½n as X ½k ¼ ½ 8 1:4 þ j1:4 k ¼ 0; 1; . . .; 7:

4 1:4 þ j1:4

8

1:4  j1:4

1:4  j1:4 ;

4

Example 3.47 For the digital signal x ½ n ¼ ½ 1

1

1

2

3 1

1

2

1 1

2

1 3

0

2

1

find 16-point DFT using decimation in time FFT algorithm. Solution 3.47 First, we divide the signal to its even and odd indexed sequences as in xe ½n ¼ ½ 1

1

x o ½ n ¼ ½ 1

2

1 1

1

1 2

3 2

1

1

3

1

0 2 :

We can calculate 8-point DFT of xe ½n and xo ½n as in the previous example. Let the calculation results be denoted by Xe ½k and Xo ½k, k ¼ 0; 1; . . .; 7. Then we can easily obtain Xe ½k and Xo ½k for k ¼ 0; 1; . . .; 15 by just repeating the elements obtained for k ¼ 0; 1; . . .; 7 and combine them using X ½k  ¼ Xe ½k  þ wk16 Xo ½k k ¼ 0; 1; . . .; 15 where the exponential vector wk16 , k ¼ 0; 1; . . .; 15 is calculated as  2p 2p ek16 ¼ ej016 ej116 2p 2p ej1016 ej1116

3.7.3

ej216 ej316 ej416 ej516 ej616 2p 2p 2p 2p ej1216 ej1316 ej1416 ej1516 : 2p

2p

2p

2p

2p

ej716

2p

ej816

2p

ej916

2p

Decimation in Frequency FFT Algorithm

Before starting the derivation of decimation in frequency FFT algorithm let’s solve some examples to become familiar with the terminology used in algorithm. Example 3.48 If x½n ¼ ½ 1

2

3

(a) Find x½n for n ¼ 0; 1; 2: (b) Find x½n for n ¼ 0; 1; . . .; 4: Solution 3.48 (a) x½n ¼ ½ 1

2

3

n ¼ 0; 1; 2

6

4

2

n ¼ 0; 1; . . .; 5:

218

3

(b) x½n ¼ ½ 1

2

3

Discrete Fourier Transform

6 4  n ¼ 0; 1; . . .; 4:

Example 3.49 If x½n ¼ ½ 1

2

3

6

2

4

n ¼ 0; 1; . . .; 5:

(a) Find x½n þ N=2 for n ¼ 0; 1; 2 and N ¼ 6. Solution 3.49 x½n þ N=2 ¼ ½ 6 4 2  n ¼ 0; 1; 2 and N ¼ 6 Note: x½nn ¼ 0; 1; . . .; N=2  1 is the first half of the signal x½n and x½n þ N=2 n ¼ 0; 1; . . .; N=2  1 is the second half of the signal x½n. Example 3.50 If x½n ¼ ½ 1

2

3

6

2 ;

4

n ¼ 0; 1; . . .; 5:

(a) Find x½n þ x½n þ N=2 for n ¼ 0; 1; 2 and N ¼ 6. (b) Find x½n  x½n þ N=2 for n ¼ 0; 1; 2 and N ¼ 6. Solution 3.50 Using the results in previous example, we obtain 

 N x½n  x n þ ¼ ½7 2

x½n þ x½n þ N=2 ¼ ½ 5

2

5

Example 3.51 For x½n ¼ ½ 2

1

3 5 ;

6

1 :

N ¼ 4, find x½nenN .

Solution 3.51 Let’s determine first enN for n ¼ 0; 1; 2; 3. Using enN ¼ ejn N the vector form of enN for n ¼ 0; 1; 2; 3 can be written as 2p

 2p enN ¼ ej0 4

ej1 4

2p

ej2 4

2p

ej3 4

2p



which can be simplified as enN ¼ ½ 1 j

1

j :

Then the product signal x½nenN for n ¼ 0; 1; 2; 3 can be written as x½nenN ¼ ½ ð2Þ  1

1  ðjÞ 3  ð1Þ 5  j 

which yields x½nenN ¼ ½ 2 j 3

j5 :

Example 3.52 X ½k  ¼ ½ 0 1 2 3 4 5 6 7  are the DFT coefficients of a digital signal x½n. Write even and odd indexed samples of X ½k as sequences. Solution 3.52 Even indexed samples are

3.7 Computation of the Discrete Fourier Transform

X ½2k ¼ ½ 0

2

219

6  k ¼ 0; 1; 2; 3

4

and odd indexed samples are X ½2k þ 1 ¼ ½ 1 3

5

7

k ¼ 0; 1; 2; 3:

Example 3.53 Even and odd indexed samples of the DFT coefficients of a digital signal are given as X ½2k ¼ ½ 1 X ½2k þ 1 ¼ ½ 2

j 2

1þj

3 1  k ¼ 0; 1; 2; 3; 4

2j 0

3  k ¼ 0; 1; 2; 3; 4

Find the DFT coefficient vector X ½k; k ¼ 0; 1; . . .; 9: Solution 3.53 Taking samples one by one from X ½2k and X½2k þ 1 in a sequential manner, we get the DFT coefficient vector X ½k ¼ ½ 1

1þj

2 j

2

2j 3

0

1

3 :

Let’s now derive the decimation in frequency FFT algorithm. Decimation in Frequency FFT Algorithm In decimation in frequency FFT algorithm the even and odd indexed DFT coefficients are calculated separately. This operation is explained as follows. The DFT coefficients are calculated using X ½k  ¼

N 1 X

x½nekn N

k ¼ 0; 1; . . .; N  1

ð3:174Þ

n¼0

from which even indexed coefficients can be obtained via X ½2k ¼

N 1 X

x½ne2kn N

k ¼ 0; 1; . . .; N=2  1

n¼0

where the summation term can be divided into two parts as X ½2k ¼

N=21 X n¼0

x½ne2kn N þ

N 1 X

x½ne2kn N

k ¼ 0; 1; . . .; N=2  1:

ð3:175Þ

n¼N=2

|fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} N 1 2 P 2kðn þ N Þ 2 x½n þ N2 eN n¼0

By changing the frontiers of the second summation expression in (3.175) we obtain

220

3

X ½2k ¼

N=21 X

x½ne2kn N

þ

N=21 X

n¼0

n¼0

  N 2kðn þ N2 Þ e x nþ 2 N

2kðn þ N2 Þ

2kðn þ N2 Þ

k ¼ 0; 1; . . .; N=2  1 ð3:176Þ

can be simplified as

where the exponential term eN eN

Discrete Fourier Transform

2kðn þ N2 Þ

kN ¼ e2kn N eN ! eN |{z}

¼ e2kn N

¼1

and making use of the e2N ¼ eN=2 the expression for X½2k in (3.176) can be written as X ½2k ¼

N=21 X

x½ne2kn N

þ

n¼0

N=21 X n¼0



 N kn x nþ e 2 N=2

k ¼ 0; 1; . . .; N=2  1

which is further simplified as X ½2k  ¼

N=21 X n¼0



N x ½ n þ x n þ 2

 k ¼ 0; 1; . . .; N=2  1:

ekn N=2

ð3:177Þ

Equation (3.177) can be written in more compact form as X ½2k ¼

N=21 X

x1 ½nekn N=2

k ¼ 0; 1; . . .; N=2  1

n¼0



where x1 ½n ¼ x½n þ x n þ N2 n ¼ 0; 1; . . .; N=2  1: In a similar manner the odd indexed coefficients of X½k can be obtained via X ½2k þ 1 ¼

N 1 X

ð2k þ 1Þn

x½neN

k ¼ 0; 1; . . .; N=2  1

n¼0

and proceeding as in the case of even indexed coefficients we obtain X ½2k þ 1 ¼

N=21 X

ð2k þ 1Þn

ðx½n  x½n þ N=2ÞeN

n¼0

which can also be written as

k ¼ 0; 1; . . .; N=2  1

3.7 Computation of the Discrete Fourier Transform

X ½2k þ 1 ¼

221

N=21 X

ðx½n  x½n þ N=2ÞenN ekn N=2

k ¼ 0; 1; . . .; N=2  1

n¼0

which can be written in more compact form as X ½2k þ 1 ¼

N=21 X

x2 ½nekn N=2

k ¼ 0; 1; . . .; N=2  1

n¼0

where x2 ½n ¼ ðx½n  x½n þ N=2ÞenN To sum it up; X ½2k ¼

N=21 X

n ¼ 0; 1; . . .; N=2  1:

x1 ½nekn N=2

k ¼ 0; 1; . . .; N=2  1

n¼0

X ½2k þ 1 ¼

N=21 X

x2 ½nekn N=2

k ¼ 0; 1; . . .; N=2  1

n¼0

where x1 ½n ¼ ðx½n þ x½n þ N=2Þ x2 ½n ¼ ðx½n  x½n þ N=2ÞenN 2p and n ¼ 0; 1; . . .; N=2  1, enN ¼ ej N n . Note: If the signal x½n is written as x½n ¼ ½ A B ; n ¼ 0; 1; . . .; N  1 where A is the first half and B is the second half of x½n, then x½n þ x½n þ N=2 ¼ ½A þ B;

n ¼ 0; 1; . . .; N=2  1

x½n  x½n þ N=2 ¼ ½A  B;

n ¼ 0; 1; . . .; N=2  1

and

and enN for n ¼ 0; 1; . . .; N=2  1 equals to  2p enN ¼ ej0 N Example 3.54 For x½n ¼ ½ 1 in frequency FFT method.

0

2

ej1 N

2p

N 2p    ej 2 N :

1  find DFT coefficients using decimation

Solution 3.54 For the given sequence and N ¼ 4 and let’s first find the signals x1 ½n and x2 ½n given as

222

3

x1 ½n ¼ ðx½n þ x½n þ N=2Þ

Discrete Fourier Transform

x2 ½n ¼ ðx½n  x½n þ N=2ÞenN

n ¼ 0; 1; . . .; N=2  1: The signal x1 ½n is obtained by summing the first and second half parts of x½n as follows 2 1  |fflfflfflffl{zfflfflfflffl} Second Half

x ½ n ¼ ½ 1 0 |fflffl{zfflffl} First Half

x1 ½n ¼ ½ 1 0  þ ½ 2 1  ! x1 ½n ¼ ½ 1 þ 2

0  1 :

To calculate x2 ½n, we first compute enN for N ¼ 4 and n ¼ 0; 1 as in h i 2p 2p en4 ¼ ej0 4 ej1 4 ! en4 ¼ ½ 1

j :

And x½n  x½n þ N=2 for N ¼ 4 is calculated by subtracting the first and second half parts of x½n as follows x½n  x½n þ 2 ¼ ½ 1

0   ½ 2 1  ! x½n  x½n þ 2 ¼ ½ 1

1 :

Thus x2 ½n is calculated as x2 ½n ¼ ðx½n  x½n þ 2Þen4 ! x2 ½n ¼ ½ 1

1 ½1

j 

which yields x2 ½n ¼ ½ 1 j : Next, we calculate the DFT coefficients of x1 ½n and x2 ½n as follows x 1 ½ n ¼ ½ 3

1  ! X1 ½k ¼ ½ 3

1

x2 ½n ¼ ½ 1 j  ! X2 ½k ¼ ½ 1

3

j 1

þ1 þj

where X1 ½k and X2 ½k  for k ¼ 0; 1 corresponds to X½2k þ 1 and X½2k respectively. Then we get  2 X ½2k þ 1 ¼ |{z} X ½1

 1  j X½2k ¼ |fflfflffl{zfflfflffl} X½0

4 |{z}



X ½3

1 þ j |fflfflffl{zfflfflffl} X½2



3.7 Computation of the Discrete Fourier Transform

223

As a result X½k becomes as X ½k ¼ ½ 1  j 2

1 þ j 4 :

Now let’s generalize this example employing parameters instead of using the numeric values. Example 3.55 For x½n ¼ ½ a frequency FFT method.

b

c

d , find DFT coefficients using decimation in

Solution 3.55 For the given sequence, let’s first find the signals x1 ½n and x2 ½n given as x1 ½n ¼ ðx½n þ x½n þ N=2Þ

x2 ½n ¼ ðx½n  x½n þ N=2ÞenN

n ¼ 0; 1; . . .; N=2  1: The signal x1 ½n is obtained by summing the first and second half parts of x½n as follows x ½ n ¼ ½ a b |fflffl{zfflffl} First Half x1 ½ n ¼ ½ a

bþ½c

c d  |fflffl{zfflffl} Second Half

d  ! x 1 ½ n ¼ ½ a þ c

b þ d :

To calculate x2 ½n, we first compute enN for N ¼ 4 and n ¼ 0; 1 as follows  2p en4 ¼ ej0 4

ej1 4

2p



! en4 ¼ ½ 1

j :

And x½n  x½n þ N=2 for N ¼ 4 is calculated by subtracting the first and second half parts of x½n as in x ½ n  x ½ n þ 2 ¼ ½ a

d  ½c

d  ! x½n  x½n þ 2 ¼ ½ a  c b  d :

Thus x2 ½n can be calculated as x2 ½n ¼ ðx½n  x½n þ 2Þen4 ! x2 ½n ¼ ½ 1

1 ½1

which yields x2 ½n ¼ ½ a  c jðb  dÞ : Next, we calculate the DFT coefficients of x1 ½n and x2 ½n, i.e.,

j 

224

3

Discrete Fourier Transform

x 1 ½ n ¼ ½ a þ c

b þ d  ! X1 ½k  ¼ ½ a þ c þ b þ d

a þ c  b  d ;

x 2 ½ n ¼ ½ a  c

jðb  dÞ  ! X2 ½k  ¼ ½ a  c  jb þ jd

a  c þ jb  jd 

where X1 ½k and X2 ½k  for k ¼ 0; 1 corresponds to X½2k þ 1 and X½2k respectively. That is  aþcþbþd X ½2k þ 1 ¼ |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl}

 a þc  b  d |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} ;

X ½1



X ½3

 a  c þ jb  jd |fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl} :

a  c  jb þ jd X ½2k  ¼ |fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl} X ½0

X ½2

As a result X½k becomes as X ½k ¼ ½ a  c  jb þ jd

aþcþbþd

a  c þ jb  jd

a þ c  b  d :

Example 3.56 For x½n ¼ ½ 2 1 1 1 3 0 1 2 , find 8-point DFT coefficients using decimation in frequency FFT method. Solution 3.56 The first and second half parts of x½n are shown in 

2 1 1 1 x½n ¼ |fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl} First Half

 3 0 1 2 |fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl} : Second Half

The signals x½n þ x½n þ N=2 x½n þ x½n þ N=2 for N ¼ 8;

enN

n ¼ 0; 1; . . .; 3 can be calculated as

x ½ n þ x ½ n þ 4 ¼ ½ 2

1  þ ½ 3

1 1

0

1 2  ! ½ 4

1 2

3 

x½n  x½n þ 4 ¼ ½ 2 1 1 1   ½ 3 0 1 2  ! ½ 1 1 0 1    2p 2p 2p 2p p p 3p en8 ¼ ej0 8 ej1 8 ej2 8 ej3 8 ! en8 ¼ 1 ej4 ej2 ej 4 : ð3:178Þ Using the results in (3.178), we can obtain the signals x1 ½n and x2 ½n as in x1 ½n ¼ ½ 4 1 2 3  x2 ½n ¼ ðx½n  x½n þ 4Þen8 x2 ½n ¼ ½ 1

1

 0 1 1

Hence we obtained the signals

ej4

p

ej2

p

ej 4

3p



!

3.7 Computation of the Discrete Fourier Transform

225

x1 ½n ¼ ½ 4 1 2 3   p 3p x2 ½n ¼ 1 ej4 0 ej 4 : The DFT coefficients of x1 ½n and x2 ½n can be found using the decimation in frequency FFT algorithm as in the previous example. Let’s denote the DFT coefficients of x1 ½n and x2 ½n as X1 ½k and X2 ½k which can be found as X1 ½k ¼ ½ 4

2  j4

8 2 þ j4 

X2 ½k  ¼ ½ 1  j2:8 1  j2:8 1 þ j2:82

1 þ j2:82 :

The Fourier coefficients of x½n, i.e., X½k are related to X1 ½k and X2 ½k via X ½2k þ 1 ¼ X1 ½k  X ½2k ¼ X2 ½k: Then we have X ½2k þ 1 ¼ ½ 4 2  j4 8 2 þ j4  X ½2k ¼ ½ 1  j2:8 1  j2:8 1 þ j2:82

1 þ j2:82 

and X½k becomes as X ½k  ¼ ½ 1  j2:8 4

3.8

1  j2:8

2  j4 1 þ j2:8

8

1 þ j2:82

2 þ j4 :

Total Computation Amount of the FFT Algorithm

Consider the calculation of the following expression x2 þ xy: Now we ask the question: How many mathematical operations are needed for the calculation of x2 þ xy ? The answer is as follows. For the computation of x2 , one multiplicative operation is needed. For the computation of xy, one multiplicative operation is needed. For the computation of x2 þ xy, two multiplicative operations and one additive operation is needed. Hence, for the computation of x2 þ xy, three mathematical operations are needed. Now consider the equality

226

3

Discrete Fourier Transform

x2 þ xy ¼ xðx þ yÞ: And we ask the same question: How many mathematical operations are needed for the calculation of xðx þ yÞ ? It is obvious that for the calculation of xðx þ yÞ; one additive operation and one multiplicative operation is needed. And the total number of mathematical operations for the calculation of xðx þ yÞ equals to two. As a result; for x2 þ xy, three mathematical operations are needed, and for xðx þ yÞ, two mathematical operations are needed. The latter one is preferable since it involves less computation amount. Decimation in time and decimation in frequency FFT algorithms are invented to decrease the computation amount for the calculation of discrete transform coefficients X½k of a digital signal x½n: We can express the total computation saving for the calculation of DFT coefficients X½k of a digital signal x½n when FFT algorithms are employed other than the direct calculation approach. For illustration purposes, in the next section, we will first calculate the total computation amount for the evaluation of DFT coefficients X½k of a digital sequence x½n. Total Computation Amount of the Direct DFT Calculation: Let’s start the discussion with an example. Example 3.57 For N ¼ 3, find the total computation amount of the DFT formula X ½k  ¼

N1 X

x½nejk N n ; 2p

k ¼ 0; 1; . . .; N  1:

n¼0

Solution 3.57 For N ¼ 3 the DFT formula takes the form X ½k  ¼

2 X

x½nejk N n ; 2p

k ¼ 0; 1; 2

n¼0

which is expanded as X ½ 0 ¼

2 X

x½nej0 3 n 2p

n¼0

X ½ 1 ¼

2 X

x½nej1 3 n 2p

n¼0

X ½ 2 ¼

2 X

x½nej2 3 n : 2p

n¼0

When the summation terms in (3.179) are expanded, we get

ð3:179Þ

3.8 Total Computation Amount of the FFT Algorithm

227

X ½0 ¼ x½0ej0 3 0 þ x½1ej1 3 0 þ x½2ej2 3 0 2p

2p

2p

X ½1 ¼ x½0ej 3 0 þ x½1ej 3 1 þ x½2ej 3 2 2p

X ½2 ¼ x½0e

2p

j22p 30

þ x½1e

j22p 31

ð3:180Þ

2p

þ x½2e

j22p 32

:

As can be seen from (3.180) for the calculation of each coefficient in (3.180), three multiplicative and two additive operations are required. Then the total number of multiplicative operations for the calculation of all the coefficients is 3  3 ¼ 9 and the total number of additive operations for the calculation of all the coefficients is 3  2 ¼ 6. In general, for the calculation of N-point DFT X½k coefficients of a digital signal x½n, N 2 multiplicative operations and N  ðN  1Þ additive operations are needed. The total computation amount is N 2 þ N  ðN  1Þ ffi 2N 2 : Now let’s consider the total computation amount of the decimation in time FFT algorithm. Total Computation Amount of the Decimation in Time FFT Algorithm Let’s solve some examples to get familiar with the expressions appearing in this section. Example 3.58 Let N ¼ 24 ; we will divide N by 2 and divide the division result by 2 also and repeat this procedure until the result equals 2. How many divisions need to be performed? Solution 3.58 24 =2 ¼ 23 ! 23 =2 ¼ 22 ! 22 =2 ¼ 2 As it is clear from the above result, 3 division operations are needed. Note: If N ¼ 2v , then v division operations are needed to get 2 at the end of successive divisions. Example 3.59 a b means that whenever you see the a term replace it by b term in a mathematical expression. Let’s define N þ2 N

N2 N2

N 2 2

if N [ 2 if N ¼ 2:

ð3:181Þ

Using (3.181), calculate the term that should be replaced for 82 . Solution 3.59 Using the definition we get 82

8 þ 2ð 4Þ 2

And for the expression 82

ð 4Þ 2

4 þ 2ð2Þ2

ð 2Þ 2

2:

8 þ 2ð4Þ2 inserting 4 þ 2ð2Þ2 for ð4Þ2 , we get

228

3

Discrete Fourier Transform

8 þ 2ð4 þ 2ð2Þ2 Þ

82

where replacing ð2Þ2 by 2, we obtain 8 þ 2ð4 þ 2  2Þ

82 which is simplified as

82

24:

ð3:182Þ

In (3.182) the obtained result equals to 8  log2 8: Note: In general; N

2

 2 N N þ2 : 2 |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl} ¼Nlog2 N

Now let’s consider the computation amount for the decimation in time FFT algorithm. In decimation in time FFT algorithm DFT coefficients X½k of x½n are calculated using X ½k ¼ G½k þ wkN H ½k

k ¼ 0; 1; . . .; N  1

ð3:183Þ

where G½k and H½k are the N=2 point DFT coefficients of even and odd indexed samples of x½n. The calculation complexities for the terms appearing on the right hand side of (3.183) can be states as:   2   N N N  1 additive operations: multiplicative and 2 2 2   2   N N N H ½k  !  1 additive operations: multiplicative and 2 2 2 G½k !

wkN H ½k ! N multiplicative operations: And lastly for the summation of G½k and wkN H ½k terms in (3.183), we need N more additive operations. Thus; the total number of multiplicative operations is  2  2 N N N2 þ þN ¼ N þ 2 2 2 which is less then N 2 , i.e.,

3.8 Total Computation Amount of the FFT Algorithm

229

N2 þ N\N 2 2 and the total number of additive operations is 0 1 0 1     N @N N @N N2  |{z} 1 Aþ  |{z} 1 AþN  N þ 2 2 2 2 2 ignore

ignore

which is less then N 2 . Hence considering the total number of multiplicative and additive operations the computational complexity is less in decimation in time FFT algorithm. Now let’s consider the number of multiplicative operations Nþ

N2 2

which can be written as  2 N N þ2 2

ð3:184Þ

2 which is replaced N for N when decimation in time FFT algorithm is applied. The term 2 in the (3.184) indicates the FFT computational complexity of G½k and H½k. If decimation in time algorithm is applied for the calculation of G½k and 2 H½k, we can replace N2 in (3.184) by

 2 N N þ2 2 4 yielding  2 !  2 N N N þ2 N þ2 ¼ N þN þ4 2 4 4 and proceeding in a similar manner and replacing

N 2 4

by

N 4

þ2

N 2 8

, we get

 2 !  2 N N N þ2 ¼ N þN þN þ8 : N þN þ4 4 8 8 This procedure is carried out until we reach to 2-point FFT calculation. If N ¼ 2v , i.e., v ¼ log2 N the successive division process results in

230

3

Discrete Fourier Transform

N þN þ  þN |fflfflfflfflfflfflfflfflfflfflfflffl ffl{zfflfflfflfflfflfflfflfflfflfflfflffl ffl} ¼ vN v terms

where replacing v by log2 N, we get N log2 N as the number of multiplicative operations required for the calculation of DFT coefficients of x½n using decimation in time FFT algorithm. A similar procedure can be carried out to find the total number of additive operations required for the calculation of DFT coefficients of x½n using decimation in time FFT algorithm.

3.9

Problems

(1) If x½n ¼ ½ 1:0

1:6 2 2:32

then   find x n2  2 ; x  n3  2 .

2:58

2:84

3 3:16

3:4 3:44

3:58

3:74

3:84

3:90 ;

x½n  2; x½n þ 3; x½n  2; x½2n  2; x½2n  2;

(2) One period of the periodic signal ~x½n around origin is x½n ¼ ½ 1 2 1 1 2 . Find one period of ~x½n  2; ~x½n þ 2; ~x½n; ~x½n  2; ~x½2n; ~x½2n; ~x½2n þ 2; ~x½2n þ 3: (3) One period of the periodic signal ~x½n around origin is x½n ¼ ½ 1 2 1 1 . Find ~x½n ~x½n:    (4) If x½n ¼ ½ 1 2 1 1 1 ; find x ðnÞ ; x ðnÞ5 ; x ð1  nÞ5 ;      5 x ð3  nÞ5 ; x ðn þ 2Þ5 ; x ðn þ 2Þ5 ; x ð2nÞ5 ; x ðn  3Þ5 ; for 0  n  4:    (5) If x½n ¼ ½ 1 2 1 1 1 ; find x ðnÞ3 ; x ð1  nÞ3 ; x ðn þ 2Þ3 ;    x ðn þ 2Þ3 ; x ð2nÞ3 ; x ðn  3Þ3 ; for 0  n  2:

Fig. 3.34 One period of the Fourier transform of the aperiodic signal x½n

X n (w) 1

0

2 1

3 2

2

w

3.9 Problems

231

(6) Calculate 4-point DFT of (7) Calculate 6-point DFT of (8) Find 5-point circular y½n ¼ ½ 1  0 3 : 1 0 |{z} 1 (9) If x½n ¼

x½n ¼ ½ 2 3 3 4 . x ¼ ½ 2 3 3 4 . convolution of x½n ¼ ½ 1 1 2

1

n¼0

(10) If x½n ¼ ½ 1

2

0 3

1 2

1 

and

 ; find

1  and

~x½n ¼

1 X

x½n  5k ;

k¼1

draw one period of the following signals. ðaÞ ~x½n

ðbÞ ~x½2  n

ðcÞ ~x½n  2

ðdÞ ~x½2n  1 :

(11) One period of the Fourier transform of the aperiodic signal x½n is shown in Fig. 3.34. (a) Find 8-point DFT of x½n i.e., X ½k ¼ ? (b) Using the DFT coefficients calculated in part (a), find x½n employing inverse DFT formula. (12) Find the convolution of x½n ¼ ½ 1 0 1 1 1 0 1 2 3 1 11 4 1 2 1 and h½n ¼ ½ 1 1 1 ] using overlap-add and overlap-save methods. (13) Find the DFT of x½n ¼ ½ 1 0 1 1 1 0 1 2  using decimation in time FFT algorithm. (14) Find the DFT of x½n ¼ ½ 1 0 1 2 1 0 1 2  using decimation in frequency FFT algorithm.

Chapter 4

Analog and Digital Filter Design

In this chapter, we will study analog and digital filter design techniques. A filter is nothing but a linear time invariant (LTI) system. Any LTI system can be described using its impulse response. If the impulse response of a LTI system is known, then for any arbitrary input the system output can be calculated by taking the convolution of the impulse response and arbitrary input. This also means that filtering operation is nothing but a convolution operation. And filter design is nothing but finding the impulse response of a linear time invariant system. For this purpose, we can work either in time domain or frequency domain. Filter systems are designed to block some input frequencies and pass others. For this reason, filter design studies are usually done in frequency domain. Fourier transform of the impulse response of the filter system is called the transfer function of the filter. To find the transfer function of filters, a number of techniques are proposed in the literature. In this chapter, we will study the most widely known techniques in the literature. Filters are divided into two main categories. These are analog filters and digital filters. In science world, more studies on analog filter design techniques are available considering the digital filter design methods. For this reason, so as to design a digital filter, usually digital filter specifications are transferred to analog domain, and analog filter design is performed then the designed analog filter is transferred to digital domain.

4.1

Review of Systems

In this chapter, we will study analog and digital filter design. Before studying filter design techniques, we will first review some fundamental concepts. We will follow the following outline in this chapter.

© Springer Nature Singapore Pte Ltd. 2018 O. Gazi, Understanding Digital Signal Processing, Springer Topics in Signal Processing 13, DOI 10.1007/978-981-10-4962-0_4

233

234

4 Analog and Digital Filter Design

Fig. 4.1 A digital system

(a) (b) (c) (d) (e) (f) (g)

x [n]

H

y [n]

Review of Systems. Review of Z-Transform. Review of Laplace Transform. Transformation between Continuous and Discrete Systems. Analogue Filter Design. IIR Digital Filter Design. FIR Digital Filter Design.

Hence, as outlined above before studying analog filter design, we will review some fundamental concepts, such as linear systems, z-transform, Laplace transform, and transformation between continuous and discrete systems. The system given in the Fig. 4.1 has input x½n and output y½n. And the relation between input and output can be indicated as y½n ¼ H fx½ng. Linearity: The system H is a linear system if for the linear combination of the inputs the system output equals to the linear combination of the individual output. This is graphically illustrated in Fig. 4.2. Mathematically the linearity property for the system H is expressed as H fax1 ½n þ bx2 ½ng ¼ aH fx1 ½ng þ bH fx2 ½ng:

ð4:1Þ

Time Invariance: The system H is time invariant if y½n  n0  ¼ Hfx½n  n0 g

ð4:2Þ

Linear and Time Invariant System: If a system is both linear and time invariant, then the system is called linear time invariant system, i.e., LTI system. For a linear time invariant system denoted by H, the impulse response is defined as h½n ¼ Hfd½ng

H

x1[n]

ax1[n] bx2 [n] Fig. 4.2 Linear system

y1[n]

x2 [ n]

H

ð4:3Þ

H

ay1[n] by2 [n]

y2 [n]

4.1 Review of Systems

235

Fig. 4.3 Impulse response and output of a linear time invariant system

[n]

H

h [n]

x [n]

H

y[ n] h[ n] x[ n]

Fig. 4.4 A LTI system

x[n]

H

n

y[ n]

x[ k ]

k

and the output of a LTI system for an arbitrary input is defined as y½n ¼ h½n  x½n

ð4:4Þ

where  denotes the convolution operation and it is evaluated as 1 X

h½ n  x ½ n ¼

h½k x½n  k:

ð4:5Þ

k¼1

This property graphically illustrated as in the following Fig. 4.3 Causality: The signal x½n is causal if x½n ¼ 0 for n\0. The linear time invariant system denoted by H is causal if h½n ¼ 0 for n\0. Difference Equations for LTI Systems: The relationship between the input and the output of a LTI system can be represented by difference equations as in N X

a½ky½n  k  ¼

k¼0

M X

b½k x½n  k

ð4:6Þ

k¼0

where y½n is the system output and x½n is the system input. Example 4.1 The system H given in Fig. 4.4 is a LTI system. (a) Write a difference equation between system input and output. (b) Determine whether the system is causal or not. Solution 4.1 (a) The relation between system input x½n and system output y½n is given as

y ½ n ¼

n X k¼1

x½k:

ð4:7Þ

236

4 Analog and Digital Filter Design

Using (4.7) then the shifted signal y½n  1 can be calculated as n1 X

y ½ n  1 ¼

x½k:

ð4:8Þ

k¼1

Taking the difference of y½n in (4.7) and y½n  1 in (4.8), we get y½n  y½n  1 ¼ x½n:

ð4:9Þ

Using (4.7) the impulse response of the system can be calculated as h½ n ¼

n X

d½k :

k¼1

ð4:10Þ

¼ u½n where it is seen that h½n ¼ 0 for n\0, which means that H is a causal system.

4.1.1

Z-Transform

For a digital sequence x½n the Z-transform is defined as X ðzÞ ¼

1 X

x½nzn

ð4:11Þ

n¼1

where the complex numbers z ¼ rejw are chosen from a circle of radius r in complex plane. Substituting z ¼ rejw into (4.11), we obtain 1 X   X rejw ¼ ðx½nr n Þejwn

ð4:12Þ

n¼1

which converges to a finite summation if 1 X

jx½nr n j\1:

ð4:13Þ

n¼1

Since z ¼ rejw then jzj ¼ r and according to (4.13) we see that the Z-transform converges only for a set of z-values and this set of z-values constitute a region in the complex plane. And this region is called region of convergence for XðzÞ.

4.1 Review of Systems

237

The Properties of the Region of Convergence: If X ðzÞ ¼ QPððzzÞÞ, then the roots of PðzÞ ¼ 0 are called the zeros of XðzÞ and the roots of QðzÞ ¼ 0 are called the poles of XðzÞ. The region of convergence of XðzÞ has the following properties. (1) The ROC does not contain any poles. (2) Fourier transform of x½n exists if the ROC of XðzÞ covers the unit circle. (3) For a right sided sequence, the ROC extends outward from the outermost pole of XðzÞ. (4) For a left sided sequence, the ROC extends inward from the innermost finite pole of XðzÞ. (5) For a finite sequence, the ROC is a ring. Example 4.2 For x½n ¼ an u½n  1, find XðzÞ. P n for the given signal, we Solution 4.2 Using the definition X ðzÞ ¼ 1 n¼1 x½nz obtain 1 X

X ðzÞ ¼

an u½n  1zn

n¼1

where u½n  1 can be replaced by  u½n  1 ¼

1 if  n  1 [ 0 ! u½n  1 ¼ 0 otherwise



1 0

leading to the calculation X ðzÞ ¼

1 X

an zn

n¼1 1 X

¼

an zn

n¼1 1 X

¼1

an zn

n¼0

1 1  a1 z 1 ¼ : 1  a1 z

¼1

 1  a z\1 ! jzj\jaj

Example 4.3 For x½n ¼ an u½n, find XðzÞ. Solution 4.3 XðzÞ ¼ 1a11 z ROC is jzj [ jaj The LTI system H given in Fig. 4.5. Can be described as in Fig. 4.6.

if n\  1 otherwise

238

4 Analog and Digital Filter Design

Fig. 4.5 LTI system

x[n ]

H

y[n ]

Fig. 4.6 LTI system with impulse response h½n

x [n]

h [n ]

y[n]

Fig. 4.7 LTI system with Z-transforms

X (z )

H (z )

Y (z )

For the system of Fig. 4.6, y½n ¼ h½n  x½n and we have Y ðzÞ ¼ HðzÞXðzÞ. The LTI system H can also be described as in Fig. 4.7 using the Z-transforms. Stability of a Discrete LTI System: For a discrete LTI system to be a stable system, its impulse response should be absolutely summable, that is: 1 X

jh½nj\1:

ð4:14Þ

n¼1

For a discrete LTI system, the transfer function is defined as H ðzÞ ¼

Y ðzÞ X ðzÞ

ð4:15Þ

And for a discrete LTI system to be a stable system, poles of HðzÞ should be inside the unit circle. Example 4.4 For a discrete LTI system, the transfer function is given as H ðzÞ ¼

z  0:5 : ðz  0:3Þðz  0:8  j0:8Þ

Determine whether the system is stable or not? Solution 4.4 The poles of HðzÞ are at z1 ¼ 0:3 and z2 ¼ 0:8 þ 0:8j, and since pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi jz2 j ¼ 0:82 þ 0:82 ! jz2 j ¼ 1:13 is outside the unit circle, the LTI system with the given transfer function is not a stable system.

4.1 Review of Systems

4.1.2

239

Laplace Transform

Laplace transform is defined for continuous time signals. The Laplace transform of hðtÞ is calculated as Z1 hðtÞest dt

H ðsÞ ¼

ð4:16Þ

1

where s is the complex frequency defined as s ¼ r þ jw. The integral expression given in (4.16) converges for some set of s values which can be represented by a region in complex plane called convergence region or region of convergence in short. The Properties of the Region of Convergence: If H ðsÞ ¼ QPððssÞÞ, then the roots of PðsÞ ¼ 0 are called the zeros of HðsÞ and the roots of QðsÞ ¼ 0 are called the poles of HðsÞ. The properties of the region of convergence (ROC) for HðsÞ can be summarized as follows. (1) (2) (3) (4) (5)

The ROC does not include any poles. The ROC consists of vertical half planes or strips. Right side signals have ROC extending in the right half plane. Left side signals have ROC extending in the left half plane. Two sided signals do either have ROC in a central vertical strip or they diverge.

Stability of a Continuous LTI Systems: The continuous LTI system H shown in Fig. 4.8. Can also be described using its impulse response as in Fig. 4.9. For the system of Fig. 4.9, we have yc ðtÞ ¼ hðtÞ  xc ðtÞ and Y ðsÞ ¼ HðsÞXðsÞ. Thus, LTI system H can also be described as using Laplace transform of the functions as in Fig. 4.10. In Fig. 4.10, HðsÞ is called the transfer function of the continuous time system.

x c (t )

H

yc (t )

hc (t )

yc (t )

Fig. 4.8 A continuous LTI system

xc (t )

Fig. 4.9 A continuous LTI system with its impulse response

X (s )

H (s )

Y (s )

Fig. 4.10 A continuous LTI system using Laplace transforms

240

4 Analog and Digital Filter Design

The continuous time system with impulse response hðtÞ is stable if its impulse response is absolutely integrable, that is, continuous LTI system is stable if Z1 jhðtÞjdt\1:

ð4:17Þ

1

If the transfer function HðsÞ of the continuous time system is known, then the stability check can be performed by inspecting the poles of HðsÞ. If all the poles of HðsÞ are in the left half plane, i.e., the complex poles have negative real parts, then the continuous time system is stable. Otherwise the system is unstable. Example 4.5 For a continuous LTI system, the transfer function is given as H ðsÞ ¼

sþ1 : ðs  0:5 þ 2jÞðs þ 3  2jÞ

Determine whether the system is stable or not? Solution 4.5 The poles of HðsÞ are s1 ¼ 0:5  2j and s2 ¼ 3 þ 2j. The system with transfer function HðsÞ is not a stable system since the pole s1 has positive real part. For continuous LTI systems, the relationship between system input and system output can be described using differential equations as in N X k¼0

4.2

a½ k 

M d k yð t Þ X d k xð t Þ ¼ b ½ k  : dtk dtk k¼0

ð4:18Þ

Transformation Between Continuous and Discrete Time Systems

We know that continuous LTI systems can be represented by differential equations. And when the continuous time system is converted to a digital system, we can represent digital system by difference equations. Now we ask the question: How can we convert a differential equation to a difference equation? For the answer of this question, let’s first inspect the conversion of dxc ðtÞ dt to its discrete equivalent. The derivative of xc ðtÞ evaluated at point t0 is nothing but the slope of the line tangent to the graph of xc ðtÞ at point t0 . This is illustrated in the Fig. 4.11.

4.2 Transformation Between Continuous and Discrete Time Systems

241

Fig. 4.11 A tangent line at point t0

xc (t )

0

t

t0

Now let’s consider the digital signal obtained from xc ðtÞ after sampling operation. The slope of the line tangent to the graph of xc ðtÞ at point t0 can be approximated using the sample values and sampling instants. The sampling of the continuous time signal is illustrated in the Fig. 4.12. The slope of the line at point t0 ¼ nTs can be calculated using the triangles as shown in the Fig. 4.13. The slope of the line tangent to the graph at point t0 ¼ nTs can be evaluated using the left triangle in Fig. 4.13 as  dxc ðtÞ xc ðnTs Þ  xc ððn  1ÞTs Þ ¼ ð4:19Þ  dt t¼nTs Ts or using the right triangle in Fig. 4.13 as  dxc ðtÞ xc ððn þ 1ÞTs Þ  xc ðnTs Þ ¼ :  dt t¼nTs Ts

ð4:20Þ

And we have the following identities x½n ¼ xc ðnTs Þ x½n  1 ¼ xc ððn  1ÞTs Þ

x½n þ 1 ¼ xc ððn þ 1ÞTs Þ:

ð4:21Þ

Using (4.21) in (4.19) and (4.20), the derivative of the continuous time signal can be written either as

Fig. 4.12 Sampling of the continuous time signal

xc ((n 1)Ts )

xc (nTs )

xc (t )

xc ((n 1)Ts )

0

( n 1)Ts

nTs

(n 1)Ts

t

242

4 Analog and Digital Filter Design

Fig. 4.13 Calculation of the slope of the tangent line at point t0 ¼ nTs

xc (( n 1)Ts )

xc ( nTs )

xc (t )

xc ((n 1)Ts )

0

or as

(n 1)Ts

nTs

(n 1)Ts

t

 dxc ðtÞ x½n  x½n  1   dt t¼nTs Ts

ð4:22Þ

 dxc ðtÞ x½n þ 1  x½n  :  dt t¼nTs Ts

ð4:23Þ

Otherwise indicated, we will use  dxc ðtÞ x½n þ 1  x½n   dt t¼nTs Ts

ð4:24Þ

for the discrete approximation of the derivative operation. In addition, the expression  dxc ðtÞ x½n þ 1  x½n   dt t¼nTs Ts is called backward difference approximation, and  dxc ðtÞ x½n  x½n  1  dt t¼nTs Ts is called forward difference approximation. Example 4.6 Obtain the discrete equivalent of the differential equation dyðtÞ þ ayðtÞ ¼ bxðtÞ: dt Solution 4.6 If the Eq. (4.25) is sampled, we obtain

ð4:25Þ

4.2 Transformation Between Continuous and Discrete Time Systems

 dyðtÞ þ ayðtÞjt¼nTs ¼ bxðtÞjt¼nTs : dt t¼nTs

243

ð4:26Þ

And substituting  dyðtÞ y½n þ 1  y½n   dt t¼nTs Ts

ð4:27Þ

y½n ¼ yðtÞjt¼nTs x½n ¼ xðtÞjt¼nTs into (4.26), we obtain the difference equation y½n þ 1  y½n þ ay½n ¼ bx½n: Ts

ð4:28Þ

If we use the forward difference approximation  dyðtÞ y½n  y½n  1  dt t¼nTs Ts we obtain y ½ n  y ½ n  1 þ ay½n ¼ bx½n Ts as the discrete approximation of (4.25). Example 4.7 Find the discrete equivalent of d 2 yð t Þ : dt2

ð4:29Þ

Solution 4.7 We can write  d 2 yðtÞ dt2 t¼nTs as  d yðtÞ dt2 



2

Substituting

¼ t¼nTs

dyðtÞ dt t¼ðn þ 1ÞT

s

Ts

  dydtðtÞ

t¼nTs

:

ð4:30Þ

244

4 Analog and Digital Filter Design

 dyðtÞ y ½ n þ 1  y ½ n  dt t¼nTs Ts into (4.30), we obtain  d 2 yðtÞ y½n þ 2  y½n þ 1  ðy½n þ 1  y½nÞ   2 dt t¼nTs Ts2 which can be simplified as  d 2 yðtÞ y½n þ 2  2y½n þ 1 þ y½n  : dt2 t¼nTs Ts2

ð4:31Þ

If we use forward difference approximation  dyðtÞ y½n  y½n  1  dt t¼nTs Ts inside the expression  d yðtÞ dt2 



2





dyðtÞ dyðtÞ dt t¼ðnT Þ  dt t¼ðn1ÞT s s

Ts

t¼nTs

we obtain  d yðtÞ dt2  2

 t¼nTs

y½ny½n1 Ts

 



y½n1y½n2 Ts

Ts

which can be simplified as  d 2 yðtÞ y½n  2y½n  1 þ y½n  2  :  2 dt t¼nTs Ts2

ð4:32Þ

Example 4.8 Find the discrete equivalent of the differential equation d 2 yðtÞ dyðtÞ dxðtÞ þ yð t Þ ¼ xð t Þ þ : þ2 dt2 dt dt Solution 4.8 If both sides of the (4.33) are sampled, we get

ð4:33Þ

4.2 Transformation Between Continuous and Discrete Time Systems

245

   d 2 yðtÞ dyðtÞ dxðtÞ þ2 þ yðtÞjt¼nTs ¼ xðtÞjt¼nTs þ : dt2 t¼nTs dt t¼nTs dt t¼nTs

ð4:34Þ

And substituting the approximations and equations  d 2 yðtÞ y½n þ 2  2y½n þ 1 þ y½n   2 dt t¼nTs Ts2  dxðtÞ x ½ n þ 1  x ½ n  dt  T s

t¼nTs

x½n ¼ xðtÞjt¼nTs y½n ¼ yðtÞjt¼nTs into (4.34), we obtain y½n þ 2  2y½n þ 1 þ y½n y ½ n þ 1  y ½ n x½n þ 1  x½n þ2 þ y½n ¼ x½n þ : Ts2 Ts Ts ð4:35Þ For Ts ¼ 1, the Eq. (4.35) reduces to y½n þ 2 ¼ x½n þ 1: Exercise: Find the discrete equivalent of d 3 yð t Þ : dt3

4.2.1

Conversion of Transfer Functions of LTI Systems

We know that continuous and discrete LTI systems can be described by differential or difference equations.

Fig. 4.14 Continuous time LTI system and its discrete equivalent

xc (t )

hc (t )

yc (t )

x[n]

h [n]

y[n]

246

4 Analog and Digital Filter Design

And a differential equation can be converted to a difference equation via sampling operation. The difference equation represents a discrete LTI system. In Fig. 4.14, a continuous time system and its discrete equivalent obtained via sampling operation is shown using block diagrams. Both continuous and discrete systems have transfer functions defined as Hc ðsÞ ¼

Yc ðsÞ Xc ðsÞ

and

Hn ðzÞ ¼

Y n ðzÞ Xn ðzÞ

respectively. Now we ask the question, given Hc ðsÞ can we obtain Hn ðzÞ from Hc ðsÞ directly? The answer to this question is yes and we will derive two methods for the direct conversion of Hc ðsÞ to Hn ðzÞ, and these methods will be called forward difference and bilinear transformation. Note: For simplicity of notation, we will drop the subscript letters c and n from the equations Hc ðsÞ and Hn ðzÞ.

4.2.2

Forward Difference Transformation Method

Consider the differential equation dyðtÞ þ ayðtÞ ¼ xðtÞ dt

ð4:36Þ

which describes a continuous LTI system. Taking the Laplace transform of both sides of (4.36), we get sY ðsÞ þ aY ðsÞ ¼ XðsÞ from which the transfer function H ðsÞ ¼ YðsÞ=XðsÞ can be calculated as H ðsÞ ¼

1 : sþa

If the differential equation dyðtÞ þ ayðtÞ ¼ xðtÞ dt

ð4:37Þ

4.2 Transformation Between Continuous and Discrete Time Systems

247

is sampled, we get  dyðtÞ þ ayðtÞjt¼nTs ¼ xðtÞjt¼nTs dt t¼nTs which yields the difference equation y ½ n  y ½ n  1 þ ay½n ¼ x½n: Ts

ð4:38Þ

And by taking the Z-transform of both sides of (4.38), we get Y ðzÞ  z1 Y ðzÞ þ aY ðzÞ ¼ XðzÞ Ts from which the transfer function HðzÞ can be calculated as H ðzÞ ¼

1 aþ

1z1 Ts

:

ð4:39Þ

When HðsÞ in (4.37) and HðzÞ in (4.39) are compared to each other as below H ðsÞ ¼

1 sþa

H ðzÞ ¼

1 aþ

1z1 Ts

we see that H ðzÞ ¼ HðsÞjs¼1z1

ð4:40Þ

Ts

Example 4.9 Obtain the discrete equivalent of d 2 yðtÞ dyðtÞ þ ayðtÞ ¼ xðtÞ þ dt2 dt

ð4:41Þ

and find the relation between HðsÞ and HðzÞ. Use forward difference transformation method. Solution 4.9 The discrete equivalent of d 2 yðtÞ dyðtÞ þ ayðtÞ ¼ xðtÞ þ dt2 dt

ð4:42Þ

248

4 Analog and Digital Filter Design

is y½n  2y½n  1 þ y½n  2 y½n  y½n  1 þ þ ay½n ¼ x½n: Ts2 Ts Laplace transform of the (4.42) is s2 Y ðsÞ þ sY ðsÞ þ aY ðsÞ ¼ X ðsÞ:

ð4:43Þ

ð4:44Þ

Z-transform difference Eq. (4.43) can be calculated as Y ðzÞ  2z1 Y ðzÞ þ z2 Y ðzÞ Y ðzÞ  z1 Y ðzÞ þ þ aY ðzÞ ¼ XðzÞ Ts2 Ts which yields

1  z1 Ts

2 Y ðzÞ 

1  z1 Y ðzÞ þ aY ðzÞ ¼ X ðzÞ: Ts

ð4:45Þ

If we compare the Laplace transform in (4.44) and Z-transform in (4.45), we see 1 that Z-transform can be obtained from Laplace transform replacing s by 1z Ts . That is ð4:46Þ

H ðzÞ ¼ HðsÞjs¼1z1 Ts

Therefore, if forward difference transformation method is used for any differential equation, the relation between transfer functions of continuous and discrete systems happens to be as in (4.46).

4.2.3

Bilinear Transformation

If the bilinear transformation method is used to obtain the difference equation from differential equation, the relation between transfer functions happens to be as H ðzÞ ¼ HðsÞj

 s¼T2s

1z1 1 þ z1



ð4:47Þ

Now let’s derive the bilinear transformation formula in (4.47). Consider the differential equation dyðtÞ þ ayðtÞ ¼ xðtÞ: dt

ð4:48Þ

4.2 Transformation Between Continuous and Discrete Time Systems

249

Let wðtÞ ¼

dyðtÞ dt

then Zt yð t Þ ¼

wðsÞds 1

which can be written as Zt0 yð t Þ ¼

Zt wðsÞds þ

1

wðsÞds t0

|fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl} yðt0 Þ

Zt yð t Þ ¼ yð t 0 Þ þ

wðsÞds:

ð4:49Þ

t0

When the Eq. (4.49) is sampled at time instants t ¼ nTs and t0 ¼ ðn  1ÞTs , we get ZnTs yðnTs Þ ¼ yððn  1ÞTs Þ þ |fflffl{zfflffl} |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl} y½n

y½n1

wðsÞds

ð4:50Þ

ðn1ÞTs

which can be written as ZnTs y½n ¼ y½n  1 þ

wðsÞds:

ð4:51Þ

ðn1ÞTs

Now let’s consider the evaluation of the integral expression in (4.51). We can evaluate the integration in (4.51) using the trapezoidal integration rule. This is shown in the Fig. 4.15. Using Fig. 4.15, we can write ZnTs wðsÞds ¼ ðn1ÞTs

Ts ðwððn  1ÞTs Þ þ wðnTs ÞÞ 2

ð4:52Þ

250

4 Analog and Digital Filter Design

w(t )

Fig. 4.15 Trapezoidal integration

w(nTs ) w ((n 1)Ts )

0

(n 1)Ts

nTs

t

which can be simplified as ZnTs wðsÞds ¼

Ts ðw½n  1 þ w½nÞ: 2

ð4:53Þ

ðn1ÞTs

Substituting (4.53) into (4.51), we obtain y ½ n ¼ y ½ n  1 þ

Ts ðw½n  1 þ w½nÞ: 2

ð4:54Þ

Consider the equation dyðtÞ þ ayðtÞ ¼ xðtÞ: dt |ffl{zffl}

ð4:55Þ

wðtÞ

When (4.55) is sampled, we obtain w½n þ ay½n ¼ x½n ! w½n ¼ ay½n þ x½n:

ð4:56Þ

If Eq. (4.56) is substituted into (4.54), we obtain y ½ n ¼ y ½ n  1 þ

Ts ðay½n  1 þ x½n  1  ay½n þ x½nÞ 2

ð4:57Þ

which can be rearranged as y ½ n þ

aTs aTs Ts Ts y ½ n þ y½n  1  y½n  1 ¼ þ x½n  1 þ x½n: 2 2 2 2

ð4:58Þ

4.2 Transformation Between Continuous and Discrete Time Systems

251

And taking the Z-transform of both sides of (4.58), we get



 aTs aTs 1 Ts  a þ z1 X ðzÞ 1þ Y ðzÞ  1  z Y ðzÞ ¼ 2 2 2

ð4:59Þ

from which the transfer function can be calculated as H ðzÞ ¼

Y ðzÞ ! H ðzÞ ¼ X ðzÞ aþ

2 Ts

1 

1z1 1 þ z1

:

ð4:60Þ

When (4.60) is compared to H ðsÞ ¼

1 aþs

ð4:61Þ

we see that H ðzÞ ¼ HðsÞj

 s¼T2s

1z1 1 þ z1



ð4:62Þ

Bilinear transformation is an efficient transformation technique. Stable continuous time LTI systems are converted into stable discrete LTI systems. That is if the poles of HðsÞ are in the left half plane, the poles of HðzÞ are inside the unit circle. This is illustrated in Fig. 4.16. Frequency Mapping in Bilinear Transformation: In bilinear transformation, the relation between continuous and digital frequency is given as s¼



2 1  z1 Ts 1 þ z1

ð4:63Þ

Digital

Analog

s-plane Fig. 4.16 Pole mapping in bilinear transformation

z-plane

252

4 Analog and Digital Filter Design

where s ¼ r þ jwa and z ¼ ejwd . Let wa ! Analog signal frequency and wd ! Digital signal frequency: Equation (4.63) yields r þ jwa ¼



2 1  ejwd Ts 1 þ ejwd wd

wd

wd

2 ej 2 ej 2  ej 2 ¼ w w w Ts ej 2d ej 2d þ ej 2d   2 sin w2d   ¼j Ts cos w2d w  2 d tan ¼j : Ts 2 Hence, wa ¼

w  2 d tan : Ts 2

!!

ð4:64Þ

Summary: Transformation of analog systems to discrete ones can be achieved by using the following methods. (1) The forwards difference transformation: s¼

z1 : Ts

(2) The backward difference transformation: s¼

z1 : Ts z

(3) The bilinear transformation:

2 1  z1 s¼ Ts 1 þ z1

4.2 Transformation Between Continuous and Discrete Time Systems

253

(4) Impulse invariance transformation: H ðzÞ ¼ Ts  Ztransform of fH ðsÞg: (5) Step invariance transformation:   H ðzÞ ¼ 1  z1  Ztransform of



H ðsÞ : s

Example 4.10 Transfer function of a continuous time system is given as H ðsÞ ¼

s2

4s þ 11 : þ 7s þ 10

Find the transfer function HðzÞ of the digital system obtained via the sampling of continuous time system.  , for simplicity of the calculation, we can Solution 4.10 H ðzÞ ¼ HðsÞj  s¼T2s

1z1 1 þ z1

choose Ts ¼ 1 and this yields H ðzÞ ¼

4.3

19 þ 22z1 þ 3z2 : 28 þ 12z1

Analogue Filter Design

Consider the continuous LTI system given in Fig. 4.17. Where the system output equals to yðtÞ ¼ xðtÞ  hðtÞ which can be written in frequency domain as Y ðwÞ ¼ X ðwÞH ðwÞ:

ð4:65Þ

If the magnitude of HðwÞ in (4.65) gets very small values for some specific values of w, the output function YðwÞ does no contain any information about XðwÞ and this operation is called filtering.

Fig. 4.17 A continuous LTI system

x(t )

h(t )

y (t )

254

4 Analog and Digital Filter Design

Any analog filter is characterized by its transfer function HðwÞ which can be a complex function with magnitude jHðwÞj and phase \HðwÞ characteristics. If we denote the phase characteristics as hðwÞ ¼ \HðwÞ ! hðwÞ ¼ argðH ðwÞÞ then phase and group delays are defined as qðwÞ ¼ 

hð w Þ dw

sðwÞ ¼ 

dhðwÞ : dw

ð4:66Þ

Group delay function gives information about the amount of delay introduced by the system transfer function to the system input. For instance, if sð w Þ ¼ 2 then for the transfer function with unit gain the system input xðtÞ ¼ sinðwtÞ yields the system output yðtÞ ¼ sinðwðt  2ÞÞ:

4.3.1

Ideal Filters

In this section we will study the transfer functions of the ideal filters. For HðwÞ, i.e., the transfer function of the ideal filter, the time domain impulse response can be calculated using the inverse Fourier transform 1 hð t Þ ¼ 2p

Z1 HðwÞdw w¼1

which is a function having non-zero values for all t values in the range 1\t\1, for this reason such filters are not physically realizable, and they are called ideal filters. Ideal Low-Pass Filter: The transfer function of the ideal low-pass filter is shown in Fig. 4.18.

4.3 Analogue Filter Design

255 H lp (w) 1

c

0

c

w

Fig. 4.18 Transfer function of the ideal low-pass filter

Whose impulse response can be calculated as 1 hlp ðtÞ ¼ 2p

Zwc 1  ejwt dw wc

1 ¼ sin cðwc tÞ pt where wc is called cut-off frequency. Ideal High-Pass Filter: The transfer function of the ideal high-pass filter is shown in Fig. 4.19. Which can be written in terms of the transfer function of the low-pass filter with the same cut-off frequency as Hhp ðwÞ ¼ 1  Hlp ðwÞ:

ð4:67Þ

whose inverse Fourier transform equals to hhp ðtÞ ¼ 1 

1 sin cðwc tÞ: pt

ð4:68Þ

Ideal Band-Pass Filter: The transfer function of the ideal band-pass filter is shown in Fig. 4.20.

H hp (w) 1

c

0

Fig. 4.19 Transfer function of the ideal high-pass filter

w c

256

4 Analog and Digital Filter Design

H bp (w)

1

ch

w

0

cl

0

cl

0

ch

Fig. 4.20 Transfer function of the ideal band-pass filter

H (w) 1

ch

w

0

cl

cl

ch

Fig. 4.21 Transfer function of the ideal band-stop filter

Which can be obtained from low-pass filter transfer function with the same cut-off frequency as Hbp ðwÞ ¼ Hlp ðw  w0 Þ þ Hlp ðw þ w0 Þ:

ð4:69Þ

In Fig. 4.20; wcl and wch are low and high cut-off frequencies. Ideal Band-Stop Filter: The transfer function of the ideal band-stop filter is shown in Fig. 4.21. Which can be obtained from band-pass filter transfer function (4.69) as Hbs ðwÞ ¼ 1  Hbp ðwÞ:

ð4:70Þ

As can be seen from the filter transfer functions; if we design a low-pass filter, we can obtain the transfer function of other filters by just manipulating the transfer function of low-pass filter. Example 4.11 The transfer function of an analog low-pass filter with cut-off frequency xc ¼ 1 rad/s is given as H 1 ðw Þ ¼

w2

1 pffiffiffi : þ 2 2w þ 4

Find the transfer function of low-pass filter with cut-off frequency xc ¼ 2 rad/s.

4.3 Analogue Filter Design

257

H 1i ( w)

Fig. 4.22 Transfer function of the ideal low-pass filter with cut-off frequency xc ¼ 1 rad/s

1

1

0

1

w

H 2i ( w)

Fig. 4.23 Transfer function of the ideal low-pass filter with cut-off frequency xc ¼ 2 rad/s

1

2

0

2

w

Solution 4.11 The transfer function of the ideal low-pass filter with cut-off frequency xc ¼ 1 rad/s is shown in the Fig. 4.21. And the transfer function of the ideal low-pass filter with cut-off frequency xc ¼ 2 rad/s is shown in the Fig. 4.4. From Figs. 4.22 and 4.23, we see that H2i ðwÞ ¼ H1i

w 2

ð4:71Þ

In a similar manner, using the low-pass filter with cut-off frequency xc ¼ 1 rad/s in the problem, we can calculate the transfer function of the low-pass filter with cut-off frequency xc ¼ 2 rad/s employing (4.71) as H 2 ðw Þ ¼

w2

4 pffiffiffi : þ 4 2w þ 16

In general, given the transfer function of low-pass filter H1 ðwÞ with cut-off frequency 1 rad/s, the transfer function of low-pass filter with cut-off frequency xc can be obtained as Hwc ðwÞ ¼ H1

w wc

ð4:72Þ

258

4 Analog and Digital Filter Design

Example 4.12 The transfer function of an analog low-pass filter with cut-off frequency xc ¼ 1 rad/s is given as H1 ðsÞ ¼

1 pffiffiffi : s2 þ 2 2 þ 4

Find the transfer function of high-pass filter with cut-off frequency xc ¼ 2 rad/s. Solution 4.12 First, we can design the low-pass filter with cut-off frequency xc ¼ 2 rad/s as in the previous example and the transfer function of the low-pass filter with cut-off frequency xc ¼ 2 rad/s is found as 4 pffiffiffi : þ 4 2w þ 16 Then the transfer function of the high-pass filter with cut-off frequency xc ¼ 2 rad/s can be found as Hlp ðwÞ ¼

w2

Hhp ¼ 1  Hlp ðwÞ pffiffiffi w2 þ 4 2w þ 12 pffiffiffi ¼ : w2 þ 4 2w þ 16 Hence, for the filter design; it is custom to design a low-pass filter with cut-off frequency xc ¼ 1 rad/s and transfer it to any desired frequency response.

4.3.2

Practical Analog Filter Design

Although ideal filters are simple to understand they cannot be used to construct filter circuits; since they need an infinite number of circuit elements. For this reason, practical analog filter design techniques are adapted in the signal processing literature. The specifications of a practical analog filter are given in Fig. 4.24. Fig. 4.24 The specifications of a practical analog filter

| H ( w) |2

1 (1

2

)

Transition

1

Passband Stopband 2

0

wp

wc

ws

w

4.3 Analogue Filter Design

259

As can be seen from Fig. 4.24, the squared filter magnitude should satisfy  1 1 þ 2  jH ðwÞj2  1

for

0  w  wp

in passband and it should satisfy 0  jH ðwÞj2  d2

for

ws  w  1

in stopband. Filter Parameters Cut-off frequency: At cut-off frequency wc , the amplitude of the transfer function equals to p1ffiffi jH ðwÞj max , that is 2 1 H ðwc Þ ¼ pffiffiffi jH ðwÞjmax : 2 If jH ðwÞjmax ¼ 1, then wc is determined from 1 H ðwc Þ ¼ pffiffiffi : 2 Pass-band ripple: Passband ripple in decibels is defined as   Rp ¼ 10 log 1 þ 2 :

ð4:73Þ

Stopband attenuation: The stopband attenuation is defined as   Rs ¼ 10 log d2 :

ð4:74Þ

Selectivity parameter: The ratio of pass-band frequency to stop-band frequency is called selectivity parameters, i.e., k¼

wp ws

which is equal to 1 for ideal filters, and for practical filters k\1.

260

4 Analog and Digital Filter Design

Discrimination parameter: The discrimination parameter is used as an indicator of the pass-band and stop-band attenuation ratios and defined as  ffi d ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 d 1 which is equal to 0 for ideal filters and d [ 1 for practical filters. Now let’s see the practical filter design methods.

4.3.3

Practical Filter Design Methods

The most known practical filter design techniques in literature are: (1) (2) (3) (4)

Butterworth filter design. Chebyshev I and II filter design. Elliptic filter design. Bessel filter design.

4.3.3.1

Butterworth Filter Design

The squared magnitude response of the Nth order Butterworth filter is defined as jH ðwÞj2 ¼ 1þ

1  2N

ð4:75Þ

w wc

where wc is the cut-off frequency. The transfer function of the Nth order Butterworth filter is H ðsÞ ¼ QN

wNc

k¼1 ðs

 pk Þ

ð4:76Þ

where the poles pk are given as jp

pk ¼ wc e 2 ð1 þ ð

2k1 N

ÞÞ :

ð4:77Þ

The transfer function HðsÞ has N poles located on a circle of radius wc on the left half plane.

4.3 Analogue Filter Design

261

Given low-pass filter specifications wp , ws ; Rp ; the low-pass Butterworth filter is designed via the following steps: (1) Using the given filter specifications and the expression jH ðwÞj2 ¼ 1þ

1  2N w wc

decide on the filter order N and cut-off frequency wc . (2) Determine the poles using pk ¼ wc e 2 ð1 þ ð jp

2k1 N

ÞÞ ;

k ¼ 1; . . .; N:

(3) Find the transfer function using the poles as

H ðsÞ ¼ QN

wNc

k¼1 ðs

 pk Þ

:

(4) And finally construct the filter circuit using the transfer function HðsÞ found in the previous step. Filter order N and cut-off frequency determination: (a) From Fig. 4.24, we see that at w ¼ wp

   H wp  ¼ 1þ

   1  2N ! H wp  ¼ wp wc

1 1 þ 2

which leads to the equation

1þ

1  2N ¼ wp wc

1 : 1 þ 2

(b) In a similar manner, from Fig. 4.24, it is also seen that at w ¼ ws

jH ðws Þj ¼ 1þ

   1 2  2N ! H wp  ¼ d ws wc

ð4:78Þ

262

4 Analog and Digital Filter Design

which yields the equation

1þ

1 2  2N ¼ d :

ð4:79Þ

wp wc

From (4.78) and (4.79), we obtain the equation set 9  2N > 2 ws = ¼ d  1 wc ! dividing them  2N wp ; ¼ 1 þ 2 > wc

we get

ws wp

2N ¼

d2  1 : 2

ð4:80Þ

When (4.80) is solved for N, we get pffiffiffiffiffiffiffiffiffiffi 3

2

d2 1

 6log   N 6 6 6 log wws p 6

7 7 7 7 7

ð4:81Þ

which can be written in terms of selectivity and discrimination parameters as &

 ’ log d1  N log 1k

ð4:82Þ

where d e is the round up to the larger integer function. And the cut-off frequency wc can be determined by solving one of the equations

ws wc

wp wc

2N ¼ d2  1

2N

ð4:83Þ ¼ 1þ

2

yielding the roots 1

wc ¼ N wp

  1 wc ¼ d2  1 2N ws :

ð4:84Þ

4.3 Analogue Filter Design

263

Or the cut-off frequency can be selected as any value from the range   1 1 N wp  wc  d2  1 2N ws :

ð4:85Þ

Example 4.13 Design the transfer function of low-pass Butterworth filter whose specifications are given as wp ¼ 1000 rad/s

ws ¼ 3000 rad/s

Rp ¼ 4 dB

Rs ¼ 40 dB:

Solution 4.13 Let’s first determine the  and d values using Rp and Rs given in the question as follows     Rp ¼ 10 log 1 þ 2 ! 4 ¼ 10 log 1 þ 2 ! 2 ¼ 1:51 !  ¼ 1:23     Rs ¼ 10 log d2 ! 40 ¼ 10 log d2 ! d2 ¼ 104 : And using the calculated 2 and d2 values in the Fig. 4.25. We can roughly sketch the filter squared magnitude response as in Fig. 4.26. Next, we determine the order N of the filter as follows w

k ¼ wps ! k ¼ 13  ffi 1:23 ffi ! d ¼ pffiffiffiffiffiffiffiffiffi d ¼ pffiffiffiffiffiffiffiffiffi ! d  0:0123 2 104 1 N

d 1 logðd1Þ logð

1 k

Þ

!N

1 logð0:0123 Þ logð3Þ

¼ 4:002 ! N ¼ 4:

And the cut-off frequency can be found using   1 1 N wp  wc  d2  1 2N ws

Fig. 4.25 Typical magnitude squared transfer function of a practical low-pass filter

| H ( w) |2

1 (1

2

)

Transition

1

Passband Stopband 2

0

wp

ws

w

264

4 Analog and Digital Filter Design

Fig. 4.26 Magnitude squared transfer function of a practical low-pass filter for Example 4.13

| H ( w) |2

1 Transition

0.4 Passband Stopband

10

4

w

0

1000

3000

as follows   1 1 1:234 1000  wc  104  1 8 3000 949:6  wc  948:69 ! wc ¼ 949 rad/s: The poles for N ¼ 4 are calculated using jp

pk ¼ wc e 2 ð1 þ ð

2k1 N

ÞÞ ;

k ¼ 1; . . .; N

as follows



j5p 5p 5p 1 þ 14Þ ð 8 p1 ¼ 949e ! p1 ¼ 949e ! p1 ¼ 949 cos þ j sin 8 8



jp j7p 3 7p 7p p2 ¼ 949e 2 ð1 þ 4Þ ! p2 ¼ 949e 8 ! p2 ¼ 949 cos þ j sin 8 8



jp j9p 5 9p 9p p3 ¼ 949e 2 ð1 þ 4Þ ! p3 ¼ 949e 8 ! p3 ¼ 949 cos þ j sin 8 8



jp j11p 11p 11p 1 þ 74Þ ð 2 8 p4 ¼ 949e ! p4 ¼ 949e ! p4 ¼ 949 cos þ j sin : 8 8 jp 2

which can be simplified as p1 ¼ 363 þ 876j p2 ¼ 876 þ 363j p3 ¼ 876  363j p4 ¼ 363  876j:

4.3 Analogue Filter Design

265

Using the calculated poles, the transfer function is evaluated as H ðsÞ ¼

wNc ð s  p1 Þ ð s  p2 Þ ð s  p3 Þ ð s  p4 Þ

which leads to the expression H ðsÞ ¼ 

9492   ðs þ 363Þ2 8762 ðs þ 876Þ2 3632

whose simplified form is H ðsÞ ¼

4.3.3.2

ðs2

900;601 : þ 726s  635;607Þðs2 þ 1752s þ 635;607Þ

Chebyshev Filter Design

Chebyshev Type-I Filter: Chebyshev Type-I filter squared magnitude response is equiripple in the passband and monotonic in the stopband. The squared magnitude response of a typical Chebyshev Type-I filter is depicted in the Fig. 4.27. In Chebyshev Type-I filter transition from passband to stopband is more rapid when compared to Butterworth filter. The square magnitude response of Chebyshev Type-I filter is defined as jHI ðwÞj2 ¼

Fig. 4.27 Square magnitude response of Chebyshev Type-I filter

1 1 þ 2 TN2

 

ð4:86Þ

w wp

| H ( w) |2

1 (1

2

)

Transition

1

Passband Stopband 2

0

wp

ws

w

266

4 Analog and Digital Filter Design

where TN ðwÞ is the Nth order Chebyshev polynomial given as  TN ðwÞ ¼

cosðN cos1 ðwÞÞ  cos h N cos h1 ðwÞ

jwj  1 : jwj [ 1

ð4:87Þ

The Chebyshev polynomial can be calculated in an iterative manner as Tm ðwÞ ¼ 2wTm1 ðwÞ  Tm2 ðwÞ

m2

ð4:88Þ

with the initial conditions T0 ðwÞ ¼ 1

and

T1 ðwÞ ¼ w:

ð4:89Þ

Chebyshev Type-I filter design: Assume that the low-pass filter specifications wp ; ws ; Rp ; Rs are given. The design of the Chebychev Type-I filter can be achieved via the following steps (1) First, with the given low-pass filter specifications; the order of the filter is determined as:  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi log d 1 þ d 2  1 cos h1 ðd 1 Þ  N pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ cos h1 ðk1 Þ log k 1 þ k2  1

ð4:90Þ

where k and d are the selectivity and discrimination parameters, and Rp is the passband ripple. The cut-off frequency is found by solving the equation Rp

H ðwc Þ ¼ 10 10 :

ð4:91Þ

(2) Next, we calculate the transfer function

H ðsÞ ¼ QN

c

k¼1 ðs

 pk Þ

ð4:92Þ

where the poles are calculated using



2k  1 2k  1 p þ jwp cos hð/Þ cos p pk ¼ wp sin hð/Þ sin 2N 2N

ð4:93Þ

4.3 Analogue Filter Design

267

in which / is defined as 1! 1 1 þ ð1 þ  2 Þ2 / ¼ ln : N 

ð4:94Þ

And the constant term c in (4.92) is calculated via 8 N Q > > pk

> : ð1 þ 2 Þ2 pk

ð4:95Þ if N is even

k¼1

Example 4.14 Design a low-pass filter whose specifications are given as wp ¼ 1000 rad/s

ws ¼ 4000 rad/s

Rp ¼ 5 dB

Rs ¼ 40 dB:

Use the transfer function of Chebyshev Type-I filter for your design. Solution 4.14 With the given filter specifications, the parameters  and d are calculated as     Rp ¼ 10 log 1 þ 2 ! 5 ¼ 10 log 1 þ 2 ! 2 ¼ 2:16 !  ¼ 1:47     Rs ¼ 10 log d2 ! 40 ¼ 10 log d2 ! d2 ¼ 104 ! d ¼ 102 : And selectivity and discrimination parameters are found via w

k ¼ wps ! k ¼ 14  ffi 1:47 ffi ! d ¼ pffiffiffiffiffiffiffiffiffi d ¼ pffiffiffiffiffiffiffiffiffi ! d  0:0147: 2 104 1 d 1

The filter order is calculated as N

cos h1 ðd 1 Þ ! N  2:38 ! N ¼ 3: cos h1 ðk1 Þ

The calculation of the poles can be achieved via 1

1 1 þ ð1 þ 2 Þ2 / ¼ ln N 

! ! / ¼ 0:2121

268

4 Analog and Digital Filter Design

pk p1 p1 p2





2k  1 2k  1 p þ jwp cos hð/Þ cos p ¼ wp sin hð/Þ sin 2N 2N p p þ j1000 cos hð0:2121Þ cos ¼ 1000 sin hð0:2121Þ sin 6 6 ¼ 106:8 þ 885:5j



3p 3p ¼ 1000 sin hð0:2121Þ sin þ j1000 cos hð0:2121Þ cos 6 6

p2 ¼ 213:7



p3 ¼ 1000 sin hð0:2121Þ sin



5p 5p þ j1000 cos hð0:2121Þ cos 6 6

p3 ¼ 106:8  885:5j: Since N is odd, the constant term is calculated using c¼

3 Y

pk ! c ¼ 170;040;000:

k¼1

Then the transfer function of the filter is calculated via H ðsÞ ¼ QN

c

k¼1 ðs

 pk Þ

leading to the expression H ðsÞ ¼

170;040;000 ðs þ 213:7Þððs þ 106:8Þ2 þ 885:52 Þ

which can be simplified as H ðsÞ ¼

170;040;000 : ðs þ 213:7Þðs2 þ 213:6s þ 784;110Þ

And the above transfer function can be implemented using operational amplifiers and passive circuit elements. Chebyshev Type-II Filter: Chebyshev Type-II filter’s magnitude squared response is monotonic in the passband and equiripple is the stopband. The magnitude squared response of a typical Chebyshev Type-II filter is depicted in the Fig. 4.28. The magnitude squared response of Type-II Chebyshev filter can be given in two different forms as

4.3 Analogue Filter Design

269

Fig. 4.28 The magnitude squared response of a typical Chebyshev Type-II filter

| H ( w) |2

1 (1

2

)

Transition

1

Passband Stopband 2

ws

wp

0

  2 TN2 wws   jHII ðwÞj ¼ 1 þ 2 TN2 wws 1  : jHI ðwÞj2 ¼ 2 1 þ  TN2 wwp

w

2

ð4:96Þ

The relationship between the two transfer functions in (4.96) is given as  2  1  jHII ðwÞj2 ¼ 1  HI w 

wp ¼

1 : ws

The transfer function of the Type-II Chebyshev filter is defined as 8 N Q > szi > > c if N is even > spi > > k¼1 > <

N Q szi H ðsÞ ¼ c > spi if N is even > > spN þ2 1 > k¼1 > > > : k 6¼ N þ 1

ð4:97Þ

ð4:98Þ

2

where zi and pi are the zeros and poles of the transfer function and they are calculated using zi ¼ j

pi ¼

cos

ws 2k1  2N

p

ð4:99Þ





ws 2k  1 2k  1 p þ j cos h ð / Þ cos p ð4:100Þ  sin h ð / Þ sin 2N 2N a2i þ b2i

270

4 Analog and Digital Filter Design

where the phase / is computed as   1 cos h1 d1 N  1  1  ¼ ln d1 þ d2  1 2 : N

/¼

ð4:101Þ

And finally the constant term c is calculated using 8 N Q > > < c¼

k¼1

pk zk

if N is even

N 1Q > > : ð1 þ 2 Þ2 pk

if N is odd:

k¼1

4.3.3.3

Elliptic Filters

The magnitude squared response of the elliptic filters are given as jH ðwÞj2 ¼

1 1 þ 2 UN2 ðwÞ

where UN ðwÞ is the Jacobian elliptic function. Elliptic filters have equiripple both in the passband and stopband. The amount of the ripple in each band can be adjusted. When the ripple in stopband approaches to zero, the filter converged to a Type-I Chebyshev filter. On the other hand, as the ripple in passband approaches to zero, the filter converged to a Type-II Chebyshev filter. If the ripples in both bands approaches to zero, then the filter converged to a Butterworth filter. Elliptic filters have the steepest roll-off characteristics. The squared magnitude response of a typical Elliptic filter is depicted in the Fig. 4.29. Fig. 4.29 The squared magnitude response of a typical Elliptic filter

| H ( w) |2

1 (1

2

)

Transition

1

Has the steepest roll-off

Passband

Stopband 2

0

wp

ws

w

4.3 Analogue Filter Design

271

The phase response of the Elliptic filters is a non-linear function. The design of the elliptic filters is relatively complex when compared to Butterworth and Chebyshev filters.

4.3.3.4

Bessel Filters

For Butterworth, Chebyshev and Elliptic filters; the group delay sðhÞ is a nonlinear function of the frequency. This means that the time delay introduced to the system varies nonlinearly with the frequency. Bessel filters are linear phase filters and the group delay for these filters is a constant number independent of the frequency. For this reason, a constant time delay is introduced into the system independent of the frequency. However, Bessel filters has the lowest roll-off factor among all the practical filters we have mentioned up to now. The squared magnitude response of a typical Bessel filter is depicted in the Fig. 4.30. Summary: Butterworth Filters: No ripple in passband and stopband. Group delay is nonlinear function of the frequency. Roll-off is low. Chebyshev Type-I Filters: Have ripple in passband, no ripple in stopband. Group delay is a nonlinear function of the frequency. Roll-off is high. Chebyshev Type-II Filters: No ripple in passband and have ripple in stopband. Group delay is nonlinear function of the frequency. Roll-off is high. Elliptic Filters: Have ripple both in passband and stopband. Group delay is a nonlinear function of the frequency. Roll-off is the highest. Bessel Filters: No ripple in passband and stopband. Group delay is constant. Roll-off is the lowest. Fig. 4.30 The squared magnitude response of a typical Bessel filter

| H ( w) |2

1 (1

2

)

Transition

1

Has the lowest roll-off

Passband

Stopband 2

0

wp

ws

w

272

4 Analog and Digital Filter Design

4.3.4

Analog Frequency Transformations

Once you have analogue low-pass prototype filter with cut-off frequency wc ¼ 1 rad/s, you can design other filters via frequency transformation. The possible frequency transformations are summarized as follows: Lowpass to lowpass

s

Lowpass to highpass s Lowpass to bandpass

s

Lowpass to bandpass

s

Lowpass to bandpass

s

s where wc is the desired cutoff frequency: wc wc where wc is the desired cutoff frequency: s s 2 þ w cl w cu : sðwcu  wcl Þ s 2 þ w cl w cu : sðwcu  wcl Þ sðwcu  wcl Þ : s2 þ wcl wcu

wcl is the lower cut-off frequence. wcu is the upper cut-off frequency. Example 4.15 The transfer function of a low-pass analog filter with cut-off frequency wc ¼ 1 rad/s is given as Hlp ðsÞ ¼

1 : ð s þ 1Þ ð s 2 þ s þ 1Þ

Using the above transfer function, find the transfer function of an high-pass analog filter with cut-off frequency wc ¼ 1 rad/s. Solution 4.15 To get the transfer function of an high-pass filter from a low-pass filter transfer function, simply replace s in low-pass filter transfer function by wsc , wc i.e., s s , that is  Hhp ðsÞ ¼ Hlp ðrÞr¼wc s

which yields the transfer function Hhp ðsÞ ¼ 1 s

þ1

1  1 s2

þ

1 s

þ1



4.3 Analogue Filter Design

273

whose simplified form can be calculated as Hhp ðsÞ ¼

s2 s : s2 þ s þ 1 s þ 1

As it is seen from the above equation, the transfer function of a high pass filter includes si like terms in the numerator.

4.4 4.4.1

Implementation of Analog Filters Low Pass Filter Circuits

Remember that the transfer function of the low-pass Butterworth filter was in the form H ðsÞ ¼ QN

wNc

k¼1 ðs

 pk Þ

:

ð4:102Þ

Considering (4.102), we can calculate the transfer function of the Butterworth filter for wc ¼ 1 and N ¼ 3 as H ðsÞ ¼

1 : ðs þ 1Þðs2 þ s þ 1Þ

ð4:103Þ

As it is also seen in (4.103), we can say that the transfer function of a low-pass filter has a constant number in its numerator, and at the denominator, we can have two different types of polynomials which are ð s þ aÞ



 s 2 þ b1 s þ b2 :

If we know how to implement ðs þ aÞ and ðs2 þ b1 s þ b2 Þ, then we can implement the transfer function HðsÞ using circuit elements. How to implement H ðsÞ ¼ a=ðs þ aÞ: The transfer function H ðsÞ ¼ a=ðs þ aÞ can be implemented using the circuit in Fig. 4.31.

Fig. 4.31 Analog implementation of H ðsÞ ¼ a=ðs þ aÞ by circuit elements

R

Vout Vin

C

274

4 Analog and Digital Filter Design

The transfer function of the circuit in Fig. 4.31 can be calculated as H ðsÞ ¼

1 Vout ðsÞ ! H ðsÞ ¼ RC 1 : Vin ðsÞ s þ RC

How to implement H ðsÞ ¼ b=ðs þ aÞ: The transfer function H ðsÞ ¼

b sþa

ð4:104Þ

can be implemented using the circuit in Fig. 4.32. The transfer function of the above circuit is H ðsÞ ¼

Vout ðsÞ ! H ðsÞ ¼ Vin ðsÞ



1 R3 R1 C 1þ R2 s þ R11C

How to implement H ðsÞ ¼ a=s2 þ b1 s þ b2 : The transfer function H ðsÞ ¼

s2

a þ b1 s þ b2

ð4:105Þ

can be implemented using the circuit in Fig. 4.33. The transfer function of the circuit in Fig. 4.33 can be calculated as H ðsÞ ¼

Vout ðsÞ  ! H ðsÞ ¼ Vin ðsÞ s2 þ s11 þ

K s1 s2 1 R 2 C1

þ

1K s2

 sþ

1 s1 s2

where K ¼ 1 þ RB =RA ; s1 ¼ R1 C1 ; s2 ¼ R2 C2 . If common values are selected for the resistors R1 ; R2 and capacitors C1 ; C2 , transfer function expression reduces to

R2

Fig. 4.32 Analog implementation of H ðsÞ ¼ b=ðs þ aÞ by circuit elements

R3

R1

Vout

Vin

C

4.4 Implementation of Analog Filters

275 C1

Fig. 4.33 Analog implementation of H ðsÞ ¼ a=s2 þ b1 s þ b2 by circuit elements

R2

R1

Vout Vin

C2 RA

RB

C1

Fig. 4.34 Alternative analog implementation of (4.105)

C2

R1

R2

Vout R3

Vin

H ðsÞ ¼ K

s2 þ

1 s2 3K 1 s s þ s2

where s ¼ RC. An alternative implementation of (4.105) can be achieved using the circuit in Fig. 4.34. The transfer function of the circuit in Fig. 4.34 can be calculated as H ðsÞ ¼

1 s1 s2

s2 þ

1 s2

sþ

1 þ R1 =R3 s1 s2

ð4:106Þ

where s1 ¼ R1 C1 ; s2 ¼ R2 C2 . If R1 and R3 are chosen as R1 ¼ R3 , then we get H ðsÞ ¼

s2 þ

1 s1 s2 1 2 s2 s þ s1 s2

:

ð4:107Þ

Example 4.16 The transfer function of second order low-pass Butterworth filter with cut-off frequency wc ¼ 1000 rad/s is given as H ðsÞ ¼

106 : s2 þ 1414s þ 2  106

Implement the given filter transfer function using circuit elements.

276

4 Analog and Digital Filter Design C1

Fig. 4.35 Second order low-pass filter implementation

C2

R1

R2

Vout Vin

R3

Solution 4.16 Let’s use the circuit given in Fig. 4.35. The transfer function of the circuit in Fig. 4.35 can be calculated as H ðsÞ ¼

1 s1 s2

s2 þ

1 s2

sþ

1 þ R1 =R3 s1 s2

:

ð4:108Þ

When (4.38) is compared to H ðsÞ ¼

106 s2 þ 1414s þ 2  106

we see that 1 ¼ 106 s1 s2

1 ¼ 1414 s2

1 þ R1 =R3 ¼ 2  106 : s1 s2

In (4.109) let’s first solve 1 ¼ 1414: s2 Since s2 ¼ R2 C2 , if C2 is chosen as 0:47 lF, then R2 ¼

1 ! R2 ¼ 1504 X: 1414  0:47  106

Next solving 1 ¼ 2  106 s1 s2

1 ¼ 1414 s2

ð4:109Þ

4.4 Implementation of Analog Filters

277

for s1 , we get s1 ¼ 1414=2  106 and if C1 is chosen as 0:47 lF, then R1 ¼ 2 

1414 ! R1 ¼ 6017 X: 0:47

Finally solving the equation 1 þ R1 =R3 ¼ 2  106 s1 s2 for 1 ¼ 106 s1 s2 and R1 ¼ 6017 X we find R3 as R3 ¼ R1 ¼ 6017 X: With the found values, our second order Butterworth low-pass filter circuit with cut-off frequency wc ¼ 1000 rad/s becomes as in Fig. 4.36. The circuit in Fig. 4.36 includes some resistor values which may not be commercially available. In this case, we should use a resistor value closest to the calculated value in the Figure. This may slightly affect the accuracy of the filter. We can use the standard resistor and capacitor values shown in Tables 4.1 and 4.2. And to get the resistor value 6017 X in our example, we can use 6:2 KX or 5:6 KX and 430 X in series.

0.47 F

Fig. 4.36 Butterworth low-pass filter circuit with cut-off frequency wc ¼ 1000 rad/s

6017

0.47 F

1504

Vout Vin

6017

278

4 Analog and Digital Filter Design

Table 4.1 Common resistor values for electronic circuits Standard resistor values (±5%) 1.0 1.1 1.2 1.3 1.5 1.6 1.8 2.0 2.2 2.4 2.7 3.0 3.3 3.6 3.9 4.3 4.7 5.1 5.6 6.2 6.8 7.5 8.2 9.1

10 11 12 13 15 16 18 20 22 24 27 30 33 36 39 43 47 51 56 62 68 75 82 91

100 110 120 130 150 160 180 200 220 240 270 300 330 360 390 430 470 510 560 620 680 750 820 910

1.0 1.1 1.2 1.3 1.5 1.6 1.8 2.0 2.2 2.4 2.7 3.0 3.3 3.6 3.9 4.3 4.7 5.1 5.6 6.2 6.8 7.5 8.2 9.1

K K K K K K K K K K K K K K K K K K K K K K K K

10 11 12 13 15 16 18 20 22 24 27 30 33 36 39 43 47 51 56 62 68 75 82 91

K K K K K K K K K K K K K K K K K K K K K K K K

100 110 120 130 150 160 180 200 220 240 270 300 330 360 390 430 470 510 560 620 680 750 820 910

K K K K K K K K K K K K K K K K K K K K K K K K

1.0 1.1 1.2 1.3 1.5 1.6 1.8 2.0 2.2 2.4 2.7 3.0 3.3 3.6 3.9 4.3 4.7 5.1 5.6 6.2 6.8 7.5 8.2 9.1

M M M M M M M M M M M M M M M M M M M M M M M M

10 11 12 13 15 16 18 20 22

M M M M M M M M M

mF mF mF mF mF mF mF mF mF mF mF mF

1.0 1.2 1.5 1.8 2.2 2.7 3.3 3.9 4.7 5.6 6.8 8.2

mF mF mF mF mF mF mF mF mF mF mF mF

10 mF

Table 4.2 Common capacitor values for electronic circuits Standard capacitor values (±10%) 10 12 15 18 22 27 33 39 47 56 68 82

pF pF pF pF pF pF pF pF pF pF pF pF

100 120 150 180 220 270 330 390 470 560 680 820

pF pF pF pF pF pF pF pF pF pF pF pF

1000 1200 1500 1800 2200 2700 3300 3900 4700 5600 6800 8200

pF pF pF pF pF pF pF pF pF pF pF pF

0.010 0.012 0.015 0.018 0.022 0.027 0.033 0.039 0.047 0.056 0.068 0.082

mF mF mF mF mF mF mF mF mF mF mF mF

0.10 0.12 0.15 0.18 0.22 0.27 0.33 0.39 0.47 0.56 0.68 0.82

22 mF 33 mF 47 lF

4.4 Implementation of Analog Filters

4.4.2

279

Analog High-Pass Filter Circuit Design

Let’s consider the transfer function of a high pass Butterworth filter given as Hhp ðsÞ ¼

s2 s : 2 s þsþ1sþ1

ð4:110Þ

Inspecting (4.110), we can conclude that the transfer function of a high pass filter contains two different terms

s2

Ks2 ; þ b1 s þ b0

as : sþb

ð4:111Þ

Then if we know how to implement the terms in (4.111) by circuit elements, then we can construct a circuit for any high pass filter. The high pass filter circuit can be obtained from a low pass filter circuit by replacing the resistors of the low pass filter by capacitors and replacing the capacitors of the low pass filter by resistors. How to implement H ðsÞ ¼ as=ðs þ bÞ: We can use the circuit in Fig. 4.37 to implement the transfer function H ðsÞ ¼

as : sþb

The transfer function of the circuit in Fig. 4.37 can be calculated in ‘s’ domain. The transfer function of the circuit in Fig. 4.37 can be calculated as H ðsÞ ¼ K

s sþ

1 R 1 C1

K ¼ 1þ

where

R2 R3

If the resistors R2 and R3 are not used in Fig. 4.37, then the transfer function reduces to H ðsÞ ¼

Fig. 4.37 Analog implementation of H ðsÞ ¼ as=ðs þ bÞ by circuit elements

s sþ

Vin

1 R 1 C1

:

ð4:112Þ

Vout

C1 R1

R3

R2

280

4 Analog and Digital Filter Design R1

Fig. 4.38 Analog implementation of H ðsÞ ¼ Ks2 =s2 þ b1 s þ b0 by circuit elements

C1

Vin

Vout

C2

R2 R3

R4

How to implement H ðsÞ ¼ Ks2 =s2 þ b1 s þ b0 : We can use the circuit in Fig. 4.38 to implement the transfer function H ðsÞ ¼

Ks2 : s 2 þ b1 s þ b0

ð4:113Þ

The circuit in Fig. 4.38 is called Sallen-Key topology whose transfer function can be calculated as H ðsÞ ¼ s2 þ



Ks2 1 s2

þ

1 R2 C 1

þ

1K s1

 sþ

1 s1 s2

ð4:114Þ

where s1 ¼ R1 C1 ; s2 ¼ R2 C2 ; K ¼ 1 þ R4 =R3 : If R1 ¼ R2 and C1 ¼ C2 , then (4.114) reduces to H ðsÞ ¼

Ks2 s2 þ 3K RC s þ

1 R2 C 2

:

ð4:115Þ

Example 4.17 Implement the high pass filter transfer function H ðsÞ ¼

2:6s2 : s2 þ 5:31s þ 176:83

ð4:116Þ

Solution 4.17 If we compare the given transfer function in (4.116) to H ðsÞ ¼

Ks2 s2 þ sþ 3K RC

1 R2 C 2

we see that 2:6s2 Ks2 ¼ s2 þ 5:31s þ 176:83 s2 þ 3K RC s þ

1 R2 C 2

4.4 Implementation of Analog Filters

281

Fig. 4.39 High pass filter circuit for Example 4.17

16 K

Vout

0.47 F 0.47 F

Vin

16 K

16 K 10 K

where we have 1 ¼ 176:83 R2 C 2 And if we choose C ¼ 0:47 lF, then R is found as 1 R2 ð0:47  106 Þ2

¼ 176:83 ! R2 ¼

1012 ! R ¼ 16000 X: 0:472  176:83

Also we have K ¼ 2:6 since, K ¼ 1 þ R4 =R3 , we get 2:6 ¼ 1 þ

R4 R4 ! ¼ 1:6 R3 R3

Since RR43 ¼ 1:6, we can choose R4 ¼ 16 KX; R3 ¼ 10 KX. Then our high pass filter circuit becomes as in Fig. 4.39. Example 4.18 Implement high pass filter transfer function H ðsÞ ¼

s2

2:6  0:5s2 þ 5:31s þ 176:83

ð4:117Þ

using circuit elements. Solution 4.18 In (4.117); we have 0.5 factor in the numerator, for this reason we add a voltage divider circuit to the end of the circuit which is shown in shadow in Fig. 4.40.

282

4.4.3

4 Analog and Digital Filter Design

Analog Bandpass Active Filter Circuits

For the implementation of analog bandpass filters, the prototype circuit shown in Fig. 4.41 can be employed.

4.4.4

Analog Bandstop Active Filter Circuits

Bandstop filters can be implemented using the circuit shown in Fig. 4.42.

16 K 1K

Vout

0.47 F 0.47 F Vin

16 K

1K

16 K 10 K

Fig. 4.40 High pass filter circuit with voltage divider Fig. 4.41 Bandpass filter circuit

R1 R1

Vout

C2

Vin

C1

R3

R5

Fig. 4.42 Bandstop filter circuit

R4

R1 C1

C2

Vout R2

Vin

C3

R3 R4

R5

4.5 Infinite Impulse Response (IIR) Digital Filter Design (Low Pass)

4.5

283

Infinite Impulse Response (IIR) Digital Filter Design (Low Pass)

Two methods are followed for the design of infinite impulse response digital filters, i.e., IIR filters. These methods are: (1) Design an analog filter and convert it to a digital filter via sampling operation, i.e., digitize the designed analog filter to get the digital filter. (2) Design the IIR digital filter directly. We will use the first approach in this book. The steps for the design of IIR filters using analog prototypes are outlined in the Table 4.3. Example 4.19 The magnitude response of a digital filter is depicted in the Fig. 4.43. (a) By mapping the digital filter specifications to a continuous time, determine the continuous time filter specifications. (b) Determine the squared magnitude response of the continuous time filter. Solution 4.19 We will use the bilinear transformation method to find the digital filter specifications. In bilinear transformation, the relationship between analog and digital frequencies is w  2 d wa ¼ tan Ts 2 which can also be written as wd ¼ 2 tan

Table 4.3 Steps for the design of an IIR digital filter

1



Ts wa : 2

IIR digital filter design using analog prototypes (1) Determine the digital filter specifications, such as wp ; ws ; Rp ; Rs (2) Map digital filter frequency specifications to continuous time filter frequency specifications using a transformation method, for instance “bilinear transformation” (3) Design the continuous time filter according to continuous time specifications (4) Transform continuous time filter to digital filter using a transformation method, for instance “bilinear transformation” (5) Implement your digital filter by either designing a hardware using digital gates, or writing a software for digital devices which can be microprocessors, digital signal processing chips, or field programmable gate arrays (FPGA)

284

4 Analog and Digital Filter Design

Fig. 4.43 The magnitude response of a digital filter

| H d ( w) | 1

0.9

0.2

0

wp

0.4

w

ws

0.8

Since digital filter specifications are given, we should use wa ¼

w  2 d tan Ts 2

1 s, then the analog pass and stop to find the analog filter specifications. Let Ts ¼ 2000 frequencies are calculated as



0:4p ! wap ¼ 2906:2 rad/s ! wap ¼ 925:54p 2

0:8p was ¼ 4000 tan ! was ¼ 12;311 rad/s ! was ¼ 3918:7p: 2

wap ¼ 4000 tan

Then the analog filter magnitude response can be drawn as in Fig. 4.44. Fig. 4.44 Analog filter magnitude response

| H a ( w) | 1

0.9

0.2 0

wap

925.54

was

w

3918.7

4.5 Infinite Impulse Response (IIR) Digital Filter Design (Low Pass)

285

Using Fig. 4.44 the squared magnitude response of the analog filter can be found as in Fig. 4.45. Example 4.20 The magnitude response of a lowpass digital filter is depicted in Fig. 4.46. State the digital filter specifications via mathematical expressions. Solution 4.20 Since Fourier transform of the digital signals is periodic with period 2p, we can express the filter specifications for the interval p  w\p. In addition, we know that aliasing in Fourier transform of a digital signal does not occur if magnitude response has nonzero values only for the interval p  w\p. For this reason, for the digital filters, we will only consider the frequency interval p  w\p. In addition, the frequency interval 0  jwj\p=2 is accepted as the low frequency region and the frequency range p=2  jwj\p is accepted as the high frequency interval.

Fig. 4.45 Squared magnitude response of the analog filter

| H a ( w) |2 1

0.81

0.04 0

Fig. 4.46 The magnitude response of a lowpass digital filter

wap

925.54

wp

0.2

was

w

3918.7

| H d ( w) | 1

0.9

0.2 0

ws

w 0.5

286

4 Analog and Digital Filter Design

Then considering Fig. 4.46, the filter response can be expressed as 0:9  jHd ðwÞj  1 jHd ðwÞj  0:2

0  jwj  0:2p; 0:5p  jwj  p:

Example 4.21 Design the digital filter with the following specifications 0:9  jHd ðwÞj  1 jHd ðwÞj  0:2

0  jwj  0:4p; 0:8p  jwj  p:

Solution 4.21 Using the given specifications we can draw the magnitude response of the digital filter as in Fig. 4.47. For the design of our digital filter, we first convert digital filter specifications to analog filter specification using the bilinear transformation method. Since this example is a continuation of Example 4.19, we can use the converted parameters of Example 4.19. Using the results of Example 4.19, we can analog draw the analog filter squared magnitude response as in Fig. 4.48. To design the analog filter, we can use one of the available analog prototypes models. Let’s choose Butterworth filter model for our design. From the given squared magnitude response in Fig. 4.48, the parameters 2 ,  and d2 are found as 1 pffiffiffiffiffiffiffiffiffiffiffiffi ¼ 0:81 ! 2 ¼ 0:2346 !  ¼ 0:4843 1 þ 2

Fig. 4.47 Digital lowpass filter for Example 4.21

d2 ¼ 0:04:

| H d ( w) |

1

0.9

0.2 0

w p 0.4

ws 0.8

w

4.5 Infinite Impulse Response (IIR) Digital Filter Design (Low Pass) Fig. 4.48 Squared magnitude response of the analog filter obtained from digital filter specifications after bilinear transformation operation

287

| H a ( w) |2 1

0.81

0.04 0

wap

925.54

was

w

3918.7

The parameters 1=d and 1=k are calculated as follows sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi rffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 d2  1 1 25  1 1 ¼ ! ¼ ! ¼ 10:1144; 2 d  d 0:2346 d 1 ws 1 ¼ ! ¼ 4:2340: k wp k And the filter order is calculated as  log d1 logð10:1144Þ 1 ! N  N ! N  1:6 ! N ¼ 2: logð4:234Þ log k The cutoff frequency is calculated via   1 1 1 1 wp N  wc  ws d2  1 2N ! 925:54p  ð0:4843Þ2  wc  3918:7p  244 leading to 1308p  wc  8619p: And considering (4.118), we can choose wc as wc ¼

1308p þ 8619p ! wc ¼ 4963p ! wc ¼ 15;592: 2

ð4:118Þ

288

4 Analog and Digital Filter Design

In the last step, the poles are calculated using pk ¼ wc ej2ð1 þ ð p

2k1 N

ÞÞ ;

k ¼ 1; . . .; N:

For N ¼ 2, the poles are found as 3p

p 1 ¼ w c ej 4

5p

p2 ¼ wc ej 4

yielding the results





 pffiffiffi pffiffiffi 3p 3p p1 ¼ 15;592 cos þ j sin ! p1 ¼ 7796  2 þ j 2 ; 4 4



 pffiffiffi pffiffiffi 5p 5p þ j sin ! p2 ¼ 7796  2  j 2 : p2 ¼ 15;592 cos 4 4

ð4:119Þ

The transfer function is found using Ha ðsÞ ¼

wNc : ðs  p1 Þðs  p2 Þ ðs  pN Þ

ð4:120Þ

Substituting the calculated poles in (4.119) into (4.120) for N ¼ 2, we get H a ðsÞ ¼ 

15;5922 pffiffiffi pffiffiffi pffiffiffi pffiffiffi s þ 7796 2  j7796 2 s þ 7796 2 þ j7796 2

which is simplified as 15;5922 Ha ðsÞ ¼  pffiffiffi2  pffiffiffi2 s þ 7796 2 þ 7796 2 leading to the result H a ðsÞ ¼

243;110;464 : s2 þ 22;050s þ 2:43  108

We are done with the analog filter design. Since our aim was to design the digital filter, we should digitize our analog filter to find the digital filter. For this purpose, we will use bilinear transformation method. The conversion procedure is outlined as: Hd ðzÞ ¼ Ha ðsÞjs¼ 2

1z1 Ts 1 þ z1

ð4:121Þ

4.5 Infinite Impulse Response (IIR) Digital Filter Design (Low Pass)

289

1 Using Ts ¼ 2000 s in (4.121), we get

Hd ðzÞ ¼  4000

1z1 1 þ z1

2

243;110;464 :   1 8 þ 22;050 4000 11z þ 2:43  10 1 þz

ð4:122Þ

When (4.122) is simplified, we obtain H d ðzÞ ¼

243;110;464  ð1 þ 2z1 þ z2 Þ 107  ð34:72 þ 12:02z1 þ 1:6z2 Þ

which can be rearranged as Hd ðzÞ ¼

24:3 þ 48:6z1 þ 24:3z2 : 34:72 þ 12:02z1 þ 1:6z2

To implement the digital filter with the above transfer function, we need to express the filter output-input relation in time domain. This is possible using H d ðzÞ ¼

Y ðzÞ Y ðzÞ 24:3 þ 48:6z1 þ 24:3z2 ! ¼ X ðzÞ X ðzÞ 34:72 þ 12:02z1 þ 1:6z2

from which we get 34:72y½n þ 12:02y½n  1 þ 1:6y½n  2 ¼ 24:3x½n þ 48:6x½n  1 þ 24:3x½n  2 which leads to the expression y½n ¼ 0:34y½n  1  0:05y½n  2 þ 0:7x½n þ 1:4x½n  1 þ 0:7x½n  2 ð4:123Þ where x½n is the input of the digital filter and y½n is the filtered signal. And the Eq. (4.123) can be implemented using a computer program, or the filter can be implemented in other digital hardware such as microprocessors, DSP chips, FPGAs, via hardware programming languages such as assembly, VHDL, etc., or an application specific digital hardware consisting of gates and other digital devices can be specifically produced for this filter.

4.5.1

Generalized Linear Phase Systems

A LTI system is said to be a generalized linear phase system if its transfer function is of the form

290

4 Analog and Digital Filter Design

H ðwÞ ¼ Ar ðwÞejðb þ awÞ

ð4:124Þ

where Ar ðwÞ is a real function of w. Considering (4.124), the group delay is calculated as sg ðwÞ ¼ 

dhðwÞ ! sg ðwÞ ¼ a: dw

ð4:125Þ

A causal LTI system is a linear phase system if its L þ 1 point impulse response h½n satisfies h½n ¼ h½L  n 0  n  L

ð4:126Þ

where L can be an odd or even integer. And for such systems, the Fourier transform of h½n happens to be in the form wL

H ðwÞ ¼ Ar ðwÞej 2 :

4.6

ð4:127Þ

Finite Impulse Response (FIR) Digital Filter Design

In many practical applications, FIR filters are preferred over their IIR counterparts. The main advantages of FIR filter over IIR filter can be summarized as follows: (1) Most IIR filters have nonlinear phase characteristics, which creates problem for practical applications. (2) FIR filters having linear phase responses and they can be easily designed. (3) FIR filters can be implemented efficiently with affordable computational overhead. (4) Stable FIR filters can be designed in an easy manner. (5) In the literature, there exist excellent FIR filter design techniques. The main disadvantage of the FIR filters over IIR filters is that for the applications requiring narrow band transitions, i.e. steep roll-off, more arithmetic operations are required which means that more digital hardware components such as adders, multiplexers, multipliers, etc., are required. Designing FIR filter is nothing but determining the impulse response of an LTI system. The impulse response of the LTI system under concern includes a finite number of samples. If h½n denotes the impulse response of a FIR filter, then the output of the filter is written as: y ½ n ¼

M X k¼L

h½kx½n  k

4.6 Finite Impulse Response (FIR) Digital Filter Design

291

where usually L ¼ M is assumed. If h½n ¼ 0 for n\0, then the filter is said to be a causal filter. Otherwise, we have an anti-causal filter. Causal filters are practically realizable; on the other hand, anti-causal filters cannot be implemented. For this reason, anti-causal FIR filters should be transferred to causal FIR filters to enable their use in practical systems.

4.6.1

FIR Filter Design Techniques

There are basically three methods used for the design of FIR filters. These methods are (a) FIR filter design by windowing. (b) FIR filter design by frequency sampling. (c) Equiripple FIR filter design. Now let’s see the first method.

4.6.1.1

FIR Filter Design by Windowing

Design of FIR Filter in Time Domain: The frequency response of an ideal low pass digital filter is shown in the Fig. 4.49 where only one period of the frequency response around origin is depicted. And we know that Hid ðwÞ satisfies Hid ðwÞ ¼ Hid ðw þ m2pÞ. The time domain expression for the low pass digital filter can be calculated as 1 hid ½n ¼ 2p

Zwc wc

Hilp ðwÞ ejwn dw |fflfflffl{zfflfflffl} ¼1

1 sinðwc nÞ ¼ pn

n ¼ 0; 1; 2; . . .

where wc is called cut-off frequency. It is clear that hid ½n includes an infinite number of samples. And the convolutional operation cannot be realized using an

H id (w)

Fig. 4.49 The frequency response of an ideal low pass digital filter

1

c

0

c

w

292

4 Analog and Digital Filter Design

infinite number of samples. To alleviate this obstacle, we truncate the ideal filter and obtain the FIR filter as  h½ n ¼

hid ½n if jnj  L 0 otherwise

which can also be written as h½n ¼ hid ½n  w½n where w½n is the rectangular window defined as  w½n ¼

1 0

if jnj  L otherwise:

This type of design approach is straightforward. However, such a designed filter suffers from Gibbs phenomenon. In addition, since the used window is anti-causal so is the FIR filter. However, we can obtain a causal window via truncation as follows  w½n ¼

1 if 0  n  L 0 otherwise:

ð4:128Þ

To alleviate the effects of Gibbs phenomenon, windows other than the rectangular one such as, Hamming, Hanning, Bartlett, Triangular, and Blackman are used. Design of FIR Filter in Frequency Domain: Assume that HðwÞ is the frequency response of a FIR filter in a way that it minimizes the error 1 ¼ 2p

Zp jH ðwÞ  Hid ðwÞj2 dw p

where applying the Parseval’s identity, we get ¼

1 X

jh½n  hid ½nj2 !

n¼1

¼

L X n¼0

jh½n  hid ½nj2 þ

X n¼Z½0 L

jh½n  hid ½nj2 :

ð4:129Þ

4.6 Finite Impulse Response (FIR) Digital Filter Design

293

When (4.129) is equated to zero, we obtain  h½n ¼

hid ½n 0

if 0  n  L otherwise:

Properties of Windows: Let Wn ðwÞ be the frequency response of the window. The main-lobe of the window is defined as the region between the first zero crossings on the left and right sides of the origin. The width of the main-lobe of the causal rectangular window is approximated as Dw ¼

4p : Lþ1

ð4:130Þ

It is desirable to have a main lobe as narrow as possible. The width of the main-lobe controls the amount of attenuation on passband region. Side-lobes are the regions extending from first zero crossings points on either side of the origin. Side-lobes are responsible for the ripples occurring in passband and stopband. For a wide range of frequencies, pass and stop band ripples are equal to each other. For the causal rectangular window increasing the window length L, decreases the width of the main-lobe, however the areas under side-lobes stays the same which means that ripples occurs with the same amplitude but more frequently. To reduce the amount of area under ripples or to reduce the height of the ripples; we need to rub the ends of the rectangular window for a smoother transition to zero. For this purpose, we employ some commonly used windows as outlined below: Hanning Window:  w½n ¼

0:5  0:5 cos 0

2pn L

if 0  n  L otherwise

ð4:131Þ

Hamming Window:  w½n ¼

0:54  0:46 cos 0

2pn L

if 0  n  L otherwise

ð4:132Þ

Blackman Window:  w½n ¼

  4pn if 0  n  L 0:42  0:5 cos 2pn L þ 0:08 cos L 0 otherwise

ð4:133Þ

294

4 Analog and Digital Filter Design

For the Hanning, Hamming, and Blackman windows the general form can be written as  w½n ¼

a þ b cos 0

2pn L

þ c cos

4pn L

if 0  n  L otherwise

ð4:134Þ

where for Hanning window a ¼ 0:5; b ¼ 0:46; c ¼ 0, and for Blackman window a ¼ 0:42; b ¼ 0:5; c ¼ 0:08. Bartlett (Triangular) Window: 8 if 0  n  L2 < 2n L 2n ð4:135Þ w½n ¼ 2  L if L2 \n  L : 0 otherwise In Table 4.4 five different windows are compared to each other considering mainlobe width and peak sidelobe amplitude. All the windows given up to now can be approximated by the Kaiser window. Now let’s give some information about Kaiser window. Kaiser Window: The Kaiser window is defined as 8 h 1 i < I0 b 1½na2 2 a w½n ¼ ð4:136Þ if 0  n  L I0 ðbÞ : 0 otherwise where I0 ð Þ is the modified Bessel function of the first kind which is equal to 1 I 0 ð xÞ ¼ 2p

Z2p ex cos h dh

ð4:137Þ

0

and a ¼ M=2; b is the design parameter given by 8 < 0:1102ðC  8:7Þ b ¼ 0:5842ðC  21Þ0:4 þ 0:07886ðC  21Þ : 0:0

C [ 50 21  C  50 C\21

Table 4.4 Windows and their properties Window type

Mainlobe width

Peak sidelobe amplitude (dB)

Rectangular Bartlett Hanning Hamming Blackman

4 p/(2L + 1) 8 p/L 8 p/L 8 p/L 12 p/L

−13 −27 −32 −43 −58

ð4:138Þ

4.6 Finite Impulse Response (FIR) Digital Filter Design

295

where the parameter C is defined as C ¼ 20 log10 q:

ð4:139Þ

2q is the maximum ripple available in the passband. Let the transition region width be defined as Dw ¼ ws  wp . With the given filter specifications, the order of the Kaiser window is found as C8 ð4:140Þ L¼ 2:285Dw which is also the length of the FIR filter satisfying the given specifications. Example 4.22 Find the impulse response of a FIR filter whose specifications are given as wp ¼ 0:4p

ws ¼ 0:8p

q ¼ 0:01:

Solution 4.22 First we need to calculate the order of the Kaiser window given as L¼

C8 2:285Dw

where the parameters are calculated as Dw ¼ ws  wp ! Dw ¼ 0:8p  0:4p ! Dw ¼ 0:4p C ¼ 20 log10 q ! C ¼ 20 log10 0:01 ! C ¼ 40 And the length of the window is found as C8 40  8 !L¼ ! L ¼ 12 2:285Dw 2:285  0:4p Next, we calculate the design parameter b as follows L¼

b ¼ 0:5842ðC  21Þ0:4 þ 0:07886ðC  21Þ ! b ¼ 0:5842ð40  21Þ0:4 þ 0:07886ð40  21Þ ! b ¼ 3:3953 The function I0 ðbÞ can be approximated as I0 ðbÞ  1 þ

b2 b4 b6 b8 þ þ þ 2 64 2304 147;456

or we need to write a computer program for the computation of the integral expression in (4.137). Using the definition of w½n

296

4 Analog and Digital Filter Design

w½n ¼

8 h 1 i < I0 b 1½na2 2 a

:

0nL otherwise

I0 ðbÞ

0

the window elements for L ¼ 12; b ¼ 3:3953; a ¼ L=2 can be calculated as w½n ¼ ½0:15 |{z}

0:31

0:5 0:69

0:85

0:96

1 0:96

0:85 0:69 0:5

0:31

0:15:

n¼0

And the FIR filter coefficients are evaluated using h½n ¼ hid ½nw½n where ideal filter coefficients are hid ½n ¼

1 sinðwc nÞ pn

for which wc can be calculated as wc ¼

wp þ ws ! wc ¼ 0:6p: 2

Hence, ideal filter coefficients can be calculated as 0:6 hid ¼ ½|{z}

 0:09

0:30

 0:06 0:07

0

 0:05

0:03 0:02

 0:03

n¼0

0

0:03

 0:016:

Finally the FIR filter coefficients are found using h½n ¼ hid ½nw½n as h½n ¼ ½0:09 0:02 4.6.1.2

0:093 0

 0:045

0:009

 0:041

0:059

0

 0:05

0:029 0:017

 0:0024

FIR Filter Design by Frequency Sampling

Let H ðwÞ be the Fourier transform of the impulse response of the FIR filter to be designed. If we take L samples from H ðwÞ via sampling operation as in

4.6 Finite Impulse Response (FIR) Digital Filter Design

H ½k  ¼ H ðwÞjw¼k2p L

297

k ¼ 0; 1; . . .; L  1

ð4:141Þ

we obtain the DFT coefficients H ½k . Using (4.141) in inverse DFT formula h½n ¼

L1 1X 2p H ½kejk L ; L k¼0

n ¼ 0; 1; ; L  1

ð4:142Þ

we obtain the impulse response of digital FIR filter.

4.7

Problems

(1) Convert the differential equation d 2 yð t Þ dyðtÞ dxðtÞ þ 3yðtÞ ¼  xð t Þ þ4 2 dt dt dt to a difference equation via sampling operation and find the transfer function of the difference equation. (2) For a continuous time LTI system, the relation between system input and system output is given via the differential equation d 2 yð t Þ dyðtÞ d 2 xð t Þ  3yðtÞ ¼ þ2 þ 2xðtÞ: 2 dt dt dt2 Considering this LTI system: (a) Find the transfer function HðsÞ of the LTI system. Decide on whether the system has the stability property or not. (b) Convert the transfer function to its discrete equivalent, for this purpose take the sampling period as Ts ¼ 1. (3) The specifications of a low-pass analog filter are given as wp ¼ 1000 rad/san

ws ¼ 8000 rad/san

Rp ¼ 10 dB

Rs ¼ 40 dB:

(a) Find the transfer function HðsÞ of this filter. In other words, design your analog filter with the given specifications in the problem. For your design, use Butterworth, Chebyshev Type-I, and Chebyshev Type-II filter design methods separately. (b) Implement your filters using circuit elements.

298

4 Analog and Digital Filter Design

(4) The specifications of a low-pass IIR digital filter are given as wp ¼ 0:1p rad/s

ws ¼ 0:7p rad/s

Rp ¼ 10 dB

Rs ¼ 40 dB:

(a) Find the transfer function HðzÞ of this filter. Use sampling period Ts ¼ 1 in your design. (b) Using HðzÞ, write a difference equation between filter input and filter output. (5) Design the FIR digital filter whose specifications are given as wp ¼ 0:4p

ws ¼ 0:8p

q ¼ 0:01:

In your design use the windowing approach, and use Kaiser window for your design.

Bibliography

1. 2. 3. 4. 5.

Discrete-Time Signal Processing by A. V. Oppenheim and R. W. Schafer. Digital Signal Processing: Principles, Algorithms, and Applications by J. G. Proakis and D. G. Manolakis. Digital Signal Processing in Communication Systems by Marvin E. Frerking. Multirate Digital Signal Processing by R. E. Crochiere and L. R. Rabiner. Digital Signal Processing by William D. Stanley.

© Springer Nature Singapore Pte Ltd. 2018 O. Gazi, Understanding Digital Signal Processing, Springer Topics in Signal Processing 13, DOI 10.1007/978-981-10-4962-0

299

Index

A Absolute value, 187 Aliasing (spectral overlapping), 30 Aliasing case in downsampled signal, 88 Aliasing in downsampling, 80 Aliasing in time domain, 182 Alternative method to compute the periodic convolution, 166 Amount of distortion, 86 Analog bandpass active filter circuits, 282 Analog bandstop active filter circuits, 282 Analog filter design, 233 Analog frequency transformations, 272 Analog high-pass filter circuit design, 279 Analogue filter design, 234, 253 Anti-aliasing filter, 126 A periodic digital signal, 150 Approximated filter, 53, 103 Approximation of the ideal interpolation filter, 111 Approximation of the reconstruction filter, 51, 52 B Backward difference approximation, 242 Bartlett (triangular) window, 294 Bessel filter design, 260 Bilinear transformation, 248 Blackman window, 293 Butterworth filter design, 260, 265 C Causality, 235 Chebyshev I and II filter design, 260 Chebyshev type-I filter, 265 Chebyshev type-II filter, 268 Circular convolution, 188 Circular shifting, 189 Combined shifting and scaling, 148, 152

Common capacitor values, 278 Common resistor values, 278 Conjugate, 187 Continuous time processing of digital signals, 61 Continuous time signal, 2, 10 Continuous to digital conversion (C/D conversion), 126, 129 Convolution, 8 Convolution in frequency domain, 186 Convolution in time domain, 186 Convolution of aperiodic digital signals, 164, 165 Convolution using overlap-add method, 199 Cut-off frequency, 259 D Decimation, 89 Decimation in frequency, 219 Decimation in frequency FFT algorithm, 217 Decimation in time, 227 Decimation in time FFT algorithm, 207 Decimator system, 89 Design of FIR filter in frequency domain, 292 Design of FIR filter in time domain, 291 DFT coefficients, 218 Difference equations for LTI systems, 235 Digital signals, 1 Digital to continuous conversion, 59 Digital to continuous converter (D/C converter), 59, 136 Discrete approximation of the derivative operation, 242 Discrete Fourier Transform, 172 Discrimination parameter, 260 Downsampled signal, 73, 85 Downsampled signal in case of aliasing, 92 Downsampler, 74, 103 Downsampling, 72

© Springer Nature Singapore Pte Ltd. 2018 O. Gazi, Understanding Digital Signal Processing, Springer Topics in Signal Processing 13, DOI 10.1007/978-981-10-4962-0

301

302 Downsampling operation, 72, 87 Drawing the fourier transform of digital signal, 25 Duality, 186, 189 E Elliptic filter design, 260 Elliptic filters, 270, 271 Even numbered samples, 214 Expansion, 97 Exponential digital signal, 157 F Fast Fourier Transform (FFT) algorithms, 207 Filter parameters, 259 Finite Impulse Response (FIR) digital filter design, 234, 290 FIR filter design by frequency sampling, 291, 296 FIR filter design by windowing, 291 FIR Filter design techniques, 291 Forward difference approximation, 242, 244 Forward difference transformation method, 246 Fourier series representation, 5 Fourier transform, 27, 29 Fourier transform of a rectangle signal, 64 Fourier transform of digital exponential signal, 120 Fourier transform of product signal, 9, 11 Frequency domain analysis of upsampling, 99 Frequency mapping, 251 G Generalized linear phase systems, 289 H Hamming window, 293 Hanning window, 293 High pass filter transfer function, 280 I Ideal band-pass filter, 255 Ideal band-stop filter, 256 Ideal filters, 254 Ideal high-pass filter, 255 Ideal low-pass filter, 254 Ideal reconstruction filter, 113 IIR digital filter design., 234 Imaginary DFT coefficients, 187 Imaginary part DFS, 187 Implementation of analog filters, 273 Impulse function, 36 Impulse response, 233

Index Impulse response of zero order hold, 137 Impulse train, 5, 7 Impulse train signal, 13 Increasing the sampling rate by an integer factor, 97 Infinite Impulse Response (IIR) digital filter design, 283 Interpolation, 103 Interpolation filter, 112 Interpretation of the downsampling, 83 Inverse Fourier Transform, 47 K Kaiser window, 294 L Laplace transform, 239 Left shifted functions, 79 L’Hôpital’s rule, 38 Linear and time invariant system, 234 Linear approximation of the reconstruction filter, 64 Linearity, 186, 189 Linear time invariant, 233 Lower cut-off frequency, 272 Lowpass digital filter, 109 Low pass filter circuits, 273 Lowpass filtering of digital signals, 121 Low pass input signal, 26 M Manipulation of digital signals, 146 Manipulation of non-periodic digital signals, 146 Manipulation of periodic digital signals, 149 Mathematical analysis of interpolation, 107 Mathematical formulization of upsampling, 98 Matrix representation of circular convolution, 196 Matrix representation of DFT and inverse DFT, 184 Mid-level quantizer, 134 Mid-rise quantizer, 135 Modified bessel function, 294 Multirate signal processing, 71 N No aliasing, 58 Non-periodic digital signal, 170 O Odd numbered elements, 214 Odd numbered samples, 214

Index One period of the fourier transform of a digital signal, 91 One period of the fourier transform of upsampled signal, 101 Overlap add, 198 Overlapped amplitudes, 15 Overlapping, 41 Overlapping line equations, 21 Overlap save, 198 Overlap-save method, 204 P Pass-band ripple, 259 Perfect reconstruction filter, 55 Periodic digital signals, 162 Phase value, 187 Practical analog filter design, 258 Practical C/D converter, 129 Practical filter design methods, 260 Practical implementations of C/D and D/C converters, 128 Product signal, 7 Properties of the discrete fourier transform, 185 Q Quality of the reconstructed signal, 103 Quantization and coding, 134 Quantizer-coder, 129 R Real DFT coefficients, 187 Real part DFS, 187 Reconstructed from the digital samples, 37 Reconstructed signal, 50 Reconstruction filter, 48, 53, 54 Reconstruction filter impulse response, 50 Reconstruction of an analog signal from its samples, 45 Reconstruction operation, 45, 54 Rectangle pulse signal, 132 Rectangular signal, 1 Region of convergence, 237 Repeating pattern, 21 Review of laplace transform, 234 Review of signal types, 158 Review of Z-transform, 234 Rotate inside, 149 Rotate left, 149 Rotate right, 149

303 S Sample and hold, 129, 130 Sampling, 00 frequency, 4, 17, 20, 33, 37 of fourier transform, 170 of the sine signal, 3 operation, 2, 5, 6, 49 period, 2, 10 Scaling of digital signals in time domain, 147 Selectivity parameter, 259 Shifted graphs, 17 Shifted replicas, 16, 27, 42 Shifting in frequency, 186 Shifting in time, 186 Shifting of digital signals in time domain, 146 Shifting of periodic digital signals, 149, 150 Shifting of the shadowed triangles, 41 Sinc(
) function, 110 Some well known digital signals, 156 Spectral overlapping problem, 13 Stability of a continuous LTI systems, 239 Stability of a discrete LTI system, 238 Stopband attenuation, 259 Symmetry, 189 T The The The The

amount of distortion, 41 delay system, 124 meaning of the aliasing, 33 properties of the region of convergence, 239 The relationship between circular and linear convolution, 198 The repeating pattern, 150 Time invariance, 234 Total computation amount, 225 Transmission overhead, 4 Trapezoidal integration, 250 Triangle shape, 23 U Unit impulse, 156 Upper cut-off frequency, 272 Upsampler, 103 Upsampling, 71, 97 Upsampling operation, 117 Z Z-Transform, 236